MiniMax released M3.1 Flash on October 19. Shortly after the release, we ran the model through the latest version of Questflow’s financial benchmark.

The result was materially different from what we saw with the previous MiniMax generation.

On our earlier benchmark run, MiniMax M3 remained clearly behind the strongest frontier models. In the new evaluation, M3.1 Flash improved by nearly 20 points and moved much closer to Claude Opus 5.5, one of the strongest models we have tested in financial workflows.

The significance of this result is not that MiniMax has suddenly become the “best financial model.” A single domain benchmark cannot support that conclusion, and broader evaluations still show meaningful differences between the two models. Artificial Analysis, for example, currently places Claude Opus 5.5 well ahead of M3.1 Flash on its aggregate Intelligence Index and on several general reasoning and terminal-based tasks. At the same time, M3.1 Flash performs competitively or better on selected agentic automation evaluations.

What is more interesting is the speed at which the gap changed.

For teams building financial Agents, this is becoming a recurring pattern. Capability that appeared concentrated among a small number of expensive frontier models can become widely available within one or two model generations. The practical frontier is therefore moving faster than conventional model rankings suggest, particularly in domains where performance depends on a combination of reasoning, structured data access, tool use, calculation and workflow execution.

Financial AI is entering a period in which capability gaps may compress much faster than product teams expect.

That changes both how models should be evaluated and how financial-agent systems should be designed.

Financial Performance Is Becoming More Domain-Specific

The AI industry still tends to discuss models as if intelligence could be represented by a single ordered list.

Claude is stronger than MiniMax. One GPT generation is stronger than another. A model that ranks higher on a broad benchmark is implicitly treated as the better choice for most applications.

That abstraction is useful at a high level, but it becomes increasingly inadequate once models are deployed inside specialized operational environments.

MiniMax M3.1 Flash illustrates the problem clearly.

Artificial Analysis gives Claude Opus 5.5 a substantial advantage on its aggregate Intelligence Index. Depending on the Opus configuration, Claude also remains significantly ahead on Terminal-Bench and several general reasoning evaluations. Yet on AutomationBench-AA, M3.1 Flash is highly competitive and in some configurations scores above Opus.

MiniMax’s own release materials report the same specialization pattern. The company highlights a score of 76.2 on BankerToolBench, which evaluates multi-step banking workflows involving document consistency, financial compliance and chained tool calls, alongside substantial gains in coding, computer use and other agentic tasks.

These differences matter because financial Agents do not operate on a generic intelligence leaderboard. They operate inside concrete workflows.

A financial Agent may need to retrieve historical financial statements, calculate derived metrics, screen thousands of securities, identify market events, rank candidates, interpret a chart, analyze wallet behavior, or execute a multi-step research process across several tools. Each of these tasks places different demands on the underlying model.

A model that is weaker in open-ended scientific reasoning can still be highly effective at structured financial screening. A model that is excellent at long-horizon coding may not be the most cost-effective choice for continuous market monitoring. Another model may perform particularly well when tasks require repetitive tool calls and structured transformations.

The relevant unit of comparison is therefore shifting from general model intelligence toward task-specific performance inside a defined financial environment.

For practitioners, the question is no longer simply:

Which model is strongest?

A more useful question is:

Which model performs this financial workflow most reliably, with these tools, at this latency and this cost?

That distinction will become increasingly important as Financial Agents mature.

A Financial Agent Is More Than Its Base Model

Financial benchmarking makes another limitation of general model comparisons particularly visible: the model itself is only one component of the system.

A production financial Agent typically operates inside an environment containing market data, fundamental data, financial definitions, code execution, search tools, specialized Skills, memory, portfolio state and permission controls.

The apparent intelligence of the Agent depends heavily on this environment.

Consider a simple example. Two models may both understand how to calculate a 20-day relative volume measure. One model retrieves the correct market history, excludes the current session from the baseline when required, calculates the ratio correctly and returns the appropriate securities. The second uses stale data or applies the wrong lookback definition.

The first system succeeds and the second fails, even though both models “know” the formula.

The difference is operational, not conceptual.

This becomes more pronounced as tasks grow more complex.

Suppose an investor asks an Agent to find U.S. equities that satisfy several conditions simultaneously: sufficient market capitalization and liquidity, volatility contraction over the previous month, price above the 50-day moving average, a breakout above a recent range on expanding volume, and no excessive move during the previous five sessions. The investor then wants the resulting securities ranked using a composite score.

The challenge is not simply understanding each indicator in isolation.

The Agent has to correctly define the universe, obtain the appropriate historical data, calculate multiple derived features, preserve the correct time ordering, apply all conditions consistently, evaluate the full universe rather than return a few plausible examples, and rank the surviving candidates correctly.

A strong language model can still fail because the surrounding data or tool architecture is weak.

Conversely, a smaller model can perform extremely well when the environment reduces ambiguity and provides reliable financial primitives.

This is one reason we increasingly view Financial AI as a systems problem rather than a pure model problem.

Why the Nearly 20-Point Improvement Matters

/home/timagent/Pictures/screenshot-2026-10-21_10-05-00.png

A large generational jump is useful because it reveals where the bottleneck may have moved.

When MiniMax M3 scored materially below leading models, one interpretation was straightforward: the underlying model capability was not yet sufficient for some classes of financial work.

When the next generation closes much of that gap, the limiting factor changes.

The bottleneck may become task design, tool quality, data quality, model routing or domain-specific evaluation.

This pattern is consistent with what we have observed more broadly in our benchmark work.

Earlier this year, several categories of financial tasks produced substantial separation among models. Scores in the 60s or 70s were not unusual. The same kinds of tasks are now increasingly solved with scores in the 90s by stronger models.

This does not mean financial AI is solved.

It means some parts of the benchmark are becoming saturated.

That distinction is critical.

A benchmark is valuable only if it continues to measure unresolved capability. Once most frontier models can perform a task reliably, the task still has value as a baseline competency test, but it contributes less information about the frontier.

The correct response is not simply to add more examples of the same task.

If every strong model can calculate revenue growth correctly, creating another 1,000 revenue-growth questions does not produce a better frontier benchmark.

The evaluation has to move toward deeper combinations of financial capabilities.

The Benchmark Must Evolve With the Models

This creates a structural problem for financial benchmarking: the question bank cannot be treated as a static asset.

Traditional benchmarks are often constructed once and then used for years. That works reasonably well when the underlying capability evolves slowly.

Foundation models are improving too quickly for that assumption.

A financial benchmark built around tasks that were challenging six months ago may become substantially less discriminative today. If the benchmark remains unchanged, scores converge toward the top and the leaderboard gradually measures historical difficulty rather than current capability.

Our experience suggests that the solution is not arbitrary difficulty inflation.

A benchmark should not become harder by hiding facts in more documents, introducing obscure terminology or adding unnecessary reasoning steps. That approach may reduce model scores without making the evaluation more relevant.

The better source of difficulty is real financial work.

Real users naturally generate difficult problems because financial decisions are compositional.

They do not usually ask for one isolated indicator. They describe an investment process.

A trader might ask an Agent to identify assets that satisfy a liquidity requirement, a particular volatility regime, a momentum condition and a breakout rule, while excluding securities that have already moved too far. A crypto investor might combine market structure, token liquidity, wallet behavior and capital flows. A portfolio manager may require a system to interpret several signals differently depending on the prevailing market regime.

These questions are difficult because multiple pieces of financial logic have to remain correct simultaneously.

That is precisely the type of difficulty a financial benchmark should preserve.

From Financial Questions to Financial Algorithms

As user requirements become more sophisticated, the boundary between a natural-language request and an algorithm begins to disappear.

Many serious financial prompts are effectively programs written in financial language.

An investor might define a process such as:

Start with all sufficiently liquid securities. Measure short- and medium-term volatility. Identify names experiencing volatility contraction. Require the price to remain above a trend measure. Detect a breakout from a recent range. Exclude securities with excessive recent appreciation. Rank the remaining candidates using a weighted combination of volume expansion, breakout strength and volatility compression.

Each individual operation is relatively simple.

The difficulty is in execution fidelity.

The model has to translate the investor’s intent into a deterministic process without silently changing definitions. It needs to understand which periods to use, whether the current bar is included, which price adjustments apply, how missing data should be treated and how ranking should be calculated.

For trading-oriented Financial Agents, this capability matters far more than the ability to generate a polished market narrative.

A model can write an excellent explanation of momentum trading and still fail to identify the correct momentum candidates.

The latter failure is what matters in production.

This is why our benchmark direction increasingly emphasizes workflows such as retrieval, calculation, market screening, signal detection, ranking and expert pattern recognition rather than financial knowledge alone.

The core question is whether a model can turn financial intent into a correct work product.

Cost Changes the Meaning of “Best”

The MiniMax result becomes more consequential when capability is considered alongside cost.

MiniMax prices M3.1 Flash at $0.15 per million input tokens and $0.60 per million output tokens. The company also states that agentic-task usage costs are approximately 70% lower than the previous generation.

Artificial Analysis lists Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens. On its standardized task-cost estimates, the gap between the two models remains substantial, although actual production cost will depend on reasoning settings, token usage, caching and tool behavior.

This cost difference matters much more in Financial Agents than in occasional chatbot usage.

A financial research application might invoke a model once to produce an investment memo. In that context, several dollars of additional inference cost may be irrelevant.

A market-monitoring Agent operates under very different economics.

It may evaluate hundreds or thousands of securities repeatedly. It may run continuously across several markets. It may perform multiple screening passes before escalating only a small subset of cases for deeper analysis.

If a lower-cost model can perform the first 90% of that workflow with comparable reliability, routing every task to the most expensive frontier model becomes economically inefficient.

The practical optimization target therefore changes.

It is not maximum benchmark score in isolation.

It is some combination of accuracy, false-positive rate, false-negative rate, latency and cost.

For many Agent architectures, this will lead naturally to model routing.

A fast and inexpensive model can monitor large universes and perform routine screening. A more expensive reasoning model can be invoked when a task becomes ambiguous, when conflicting evidence appears or when a high-value decision requires deeper analysis.

The architecture starts to resemble a research organization more than a chatbot.

Not every task needs the most senior analyst.

Model Routing May Matter More Than Model Loyalty

This leads to an important engineering implication.

Financial-agent products should probably avoid becoming structurally dependent on one model provider.

Model leadership is becoming too temporary.

A workflow that requires Claude today may be viable on MiniMax tomorrow. A new Gemini, GPT, DeepSeek or Kimi release may change the trade-off again several weeks later.

The system therefore benefits from separating the financial environment from the model.

Data access should remain consistent.

Tool schemas should remain stable.

Financial metric definitions should remain versioned.

Evaluation should be reproducible.

The underlying model can then be replaced or routed according to measured performance.

This is also why benchmarking has practical value beyond publishing a leaderboard.

A benchmark becomes part of model infrastructure.

When a new model is released, it can be inserted into the same financial environment and evaluated against the same classes of work. Teams can then answer concrete questions: Is it reliable enough for screening? Does it miss more qualifying assets? Is ranking quality better? Does the cost advantage justify a small accuracy trade-off? Which task families should be routed to it?

That is much more useful than knowing that a model gained several points on a general intelligence index.

The Durable Advantage Moves Outside the Model

As model capability becomes more abundant, the source of durable differentiation changes.

Access to a strong foundation model is no longer rare. Most teams can call the same APIs.

What remains difficult to reproduce is the domain system surrounding the model.

In finance, that includes reliable historical and real-time data, canonical metric definitions, high-quality tools, specialized Skills, portfolio context, persistent memory, deterministic evaluation, execution infrastructure and risk controls.

It also includes something even more important: financial judgment.

A model may become extremely good at following a sophisticated trading framework. That does not mean the model created the framework.

This distinction matters for how we think about Financial Agents at Questflow.

We do not assume that the primary source of differentiated investment judgment should be a generic model generating strategies from scratch. Professional investors and portfolio managers already possess differentiated ways of interpreting markets. They decide which signals matter, how signals interact, what invalidates a thesis, how risk should be sized and when a technically valid setup should still be ignored.

AI can make that judgment more scalable.

It can convert parts of an investor’s process into structured rules and monitoring tasks. It can watch the market continuously, identify conditions that match the framework, explain what changed and eventually execute within explicit permissions.

As foundation models become cheaper and more capable, the economics of scaling this judgment improve dramatically.

This may be one of the most important consequences of capability compression.

The value is not simply that existing AI products become cheaper.

Entire categories of persistent financial analysis become economically feasible.

Financial Intelligence Is Moving From Knowledge to Execution Fidelity

The progression of financial AI can be viewed in several stages.

The first stage emphasized knowledge: does the model understand financial concepts?

The second emphasized analysis: can it retrieve information, calculate metrics and reason over financial documents?

The current stage increasingly emphasizes operational fidelity: can it correctly apply a financial process over real data?

This is where the benchmark becomes much closer to the requirements of trading systems.

For a trading-oriented Agent, a plausible answer is insufficient.

The result must be correct enough to act on.

If the task is to find 20 qualifying assets and the model returns 17, the three missing assets are false negatives.

If it returns five securities that do not satisfy the conditions, those are false positives.

If it identifies the correct set but ranks them incorrectly, that is a ranking error.

These failures can be measured directly.

This is why Financial Agent evaluation should increasingly resemble a combination of information retrieval, quantitative testing and workflow validation rather than traditional language-model judging.

The closer AI gets to capital allocation, the less useful “looks reasonable” becomes as an evaluation standard.

What the MiniMax Result Actually Tells Us

The most important conclusion from our latest run is therefore not that MiniMax M3.1 Flash has overtaken one particular model.

It has not demonstrated that universally, and broader benchmarks continue to show clear areas where Claude Opus 5.5 is stronger.

The result tells us something more useful.

The gap between a relatively inexpensive model and a top-tier frontier model can shrink dramatically within one generation for a specific class of financial tasks.

That has three implications.

First, financial capability is becoming increasingly domain-specific. General model rankings are useful but insufficient for production model selection.

Second, the economics of Financial Agents are changing as quickly as their capabilities. When performance converges while inference costs remain far apart, architecture and routing become strategic decisions.

Third, financial benchmarks must evolve continuously. A task set that successfully discriminates models today may become saturated within months.

For us, this reinforces a simple principle: the benchmark cannot be the dataset.

The benchmark must be a living evaluation system connected to real financial work.

Its role is to continuously translate user requirements, market workflows and expert investment processes into reproducible tests of model capability.

As models master one class of task, the evaluation has to move toward the next unresolved layer.

That is how a benchmark remains useful in a field moving this quickly.

The Frontier Is Moving From Models to Systems

The broader direction is becoming clearer.

Foundation-model capability will continue to improve. Price-performance ratios will continue to fall. Individual model leadership will continue to change.

That makes it increasingly difficult to build a durable financial product around the proposition that it uses one particular model.

The more defensible layer is the financial system built around interchangeable models.

A system with high-quality data, reliable tools, strong evaluation, sophisticated routing and structured sources of investment judgment can benefit every time a new model improves.

Instead of being disrupted by model turnover, it absorbs it.

MiniMax M3.1 Flash is a useful example.

A nearly 20-point generational improvement changes the economics of where that model can be deployed. Tasks that may have required a more expensive model can potentially move to a lower-cost layer. Continuous screening can become cheaper. More assets can be monitored. More financial workflows can run in parallel.

The next model release may change the architecture again.

That should be expected.

For Financial AI, the durable competitive advantage will increasingly come from designing systems that assume intelligence itself will continue to become cheaper and more available.

The model remains important.

But the model is becoming one component in a much larger financial stack.

And as the capability gap compresses, the real differentiation shifts toward the quality of the data, the financial workflows, the evaluation framework, the risk architecture and, ultimately, the judgment that the Agent is designed to operationalize.

For decades, consumer finance has relied on an asymmetry that is so familiar that it is rarely treated as a structural feature of the system: financial institutions optimize continuously, while most customers do not.

Banks manage their balance sheets every day. Brokerage firms optimize cash economics, execution and funding. Asset managers monitor portfolios against risk limits. Credit desks continuously evaluate spreads, collateral and repayment conditions. The average consumer, by contrast, makes financial decisions intermittently. Cash remains in low-yield accounts, portfolios drift from their original allocations, refinancing opportunities are missed, and brokerage balances remain unused for long periods of time.

This behavior is not necessarily irrational. Continuous financial optimization has historically required too much attention relative to the economic value of many individual decisions. A person may know that a better savings rate exists elsewhere and still decide that opening another account, reviewing the conditions, transferring funds and monitoring the new account is not worth the effort. A slightly better execution price, a temporary lending-rate difference or a marginally more efficient cash allocation may be economically real, but too small to justify human attention.

At the scale of an individual household, these decisions appear minor. At the scale of the financial system, they form an enormous source of economic value. Banks benefit from deposits that remain cheap and stable. Brokers benefit when cash sits idle. Asset managers benefit when investors are slow to move. Credit providers benefit when customers do not continuously refinance. In many parts of finance, a portion of the margin exists because customers do not optimize every decision every day.

Agentic AI changes this assumption because it changes the cost of attention.

An AI Agent does not need to remember to compare interest rates. It does not become tired of monitoring a portfolio. It can evaluate thousands of alternatives continuously, apply the same rules repeatedly, and identify changes that would be too small or too frequent for a human to monitor manually. As financial Agents gain access to account data and increasingly sophisticated execution infrastructure, they can move from simply explaining opportunities to maintaining financial objectives over time.

That transition could matter far more than another generation of AI-powered market predictions.

The deeper change is that capital itself may become less passive.

Finance Has Long Monetized the Cost of Attention

Bank deposits are the clearest example of this dynamic.

A bank does not earn attractive economics from deposits merely because customers hold money there. The value comes from the fact that a meaningful portion of those deposits is both cheap and persistent. Customers may leave large balances in checking accounts paying little or no interest even when Treasury bills, money-market funds or competing savings products offer materially higher yields.

That spread exists partly because customers value convenience and partly because the cost of finding and implementing a better alternative has historically been non-trivial.

Recent analysis has highlighted just how large this economic exposure may be. The Financial Times noted that JPMorgan Chase, Bank of America and Wells Fargo together hold roughly $1.6 trillion in non-interest-bearing deposits. If a substantial portion of those balances had to be repriced toward more competitive yields, the effect on bank earnings and valuations could be significant. The point is not that all of these deposits will suddenly disappear. It is that the economic value of deposit stickiness becomes less secure when software can continuously challenge it.

This logic extends far beyond banking.

A consumer with several credit cards rarely calculates the optimal card for every transaction. A small business rarely reallocates every cash balance based on short-lived yield differences. A retail investor does not continuously compare execution quality across trading venues. A homeowner may refinance a mortgage only a handful of times over decades, even when economic conditions change much more frequently.

In each case, there is a threshold below which the benefit of optimization is too small to justify human effort.

Agents lower that threshold.

If evaluating one more alternative costs almost nothing, then smaller inefficiencies become worth correcting. A 20-basis-point yield difference that a human ignores may still be economically relevant to a system monitoring billions of dollars or millions of users. A temporary change in financing conditions may become actionable even if it lasts only a few days. An unused cash balance can be evaluated continuously rather than only when the user happens to notice it.

The important distinction is not that capital will move constantly.

It is that capital can be evaluated constantly.

That is a very different financial environment.

From Advice to Mandates

Most financial AI products today are still organized around questions and answers.

A user asks whether one savings account is better than another. The model compares them. The user decides whether to act.

A user asks whether a portfolio is too concentrated. The system produces an analysis. The user determines what to change.

This structure remains fundamentally episodic. The AI becomes active only when the user initiates the interaction.

An Agent can operate differently because it can maintain a persistent mandate.

A user may define an objective such as maintaining six months of emergency liquidity while maximizing the yield on excess cash, subject to limits on credit risk and withdrawal restrictions. Once that mandate exists, the system does not need to wait for the user to ask whether the current allocation is still appropriate. It can monitor rates, liquidity conditions and account balances continuously and surface a decision when the portfolio moves outside the intended framework.

The same principle can apply to investing.

A user might define a long-term equity allocation, a maximum acceptable drawdown, limits on concentration and a set of conditions under which risk should be reduced. The Agent can continuously compare the portfolio and market environment against those instructions.

At that point, financial AI is no longer merely answering a question. It is maintaining a financial objective.

That is the conceptual shift at the center of Agentic Finance.

A financial application has traditionally been a place where the user goes to perform an action. A financial Agent can instead become a system that continuously represents the user’s financial intent.

Once that happens, the frequency of financial decision-making changes. The user does not need to repeatedly rediscover the same problem. The Agent preserves the mandate and evaluates the financial world against it.

This is the point at which many existing financial business models begin to look different.

Customer Loyalty Becomes Less Economically Reliable

Financial institutions have historically invested heavily in owning the customer relationship because distribution is one of the most valuable assets in finance.

Branches once played this role. Later, websites and mobile applications became the primary interface. If a bank owned the customer’s checking account, it had a strong chance of selling that customer a credit card, a mortgage or an investment product. If a brokerage owned the trading interface, it also captured cash balances, order flow and adjacent financial activity.

Agentic Finance weakens the connection between owning the interface and owning the economics.

A user may continue to consider one bank their primary bank while an Agent moves excess cash to another institution. They may keep one brokerage application as their preferred interface while an Agent routes specific transactions elsewhere when execution or financing conditions are better. They may remain loyal to a familiar financial brand without leaving all of their capital economically captive to that brand.

In other words, the institution can retain the customer while losing some of the customer’s financial activity.

That distinction is strategically important.

Historically, switching costs helped turn a customer relationship into economic persistence. If software reduces those switching costs, customer ownership becomes less valuable unless the underlying product remains competitive.

The competitive advantage therefore begins to move upward in the stack.

A bank still provides a deposit product. A brokerage still provides execution. An exchange still provides liquidity. An asset manager still provides an investment vehicle. But another system may increasingly decide which of those products receives the next dollar.

That system is the Agent.

The Agent Could Become the New Distribution Layer

This suggests a broader change in financial distribution.

Financial firms have traditionally designed products for human comparison.

Branding matters because humans use shortcuts. Interface design matters because users value convenience. Familiarity matters because people are reluctant to change financial providers. Marketing matters because attention is scarce.

An Agent operates under a different set of constraints.

It can evaluate structured information more consistently. It can compare rates, fees, liquidity, tax treatment, counterparty exposure and execution quality repeatedly. It can apply the same mandate across multiple providers without becoming fatigued by the process.

This does not make branding irrelevant. Trust will remain central to finance. Institutions will still need strong reputations, reliable infrastructure and regulatory credibility.

But the product also needs to become legible to software.

A future bank may need to compete not only for a consumer’s attention, but for the recommendation of the consumer’s Agent. A trading venue may need to compete not only for active users, but for Agent-directed order flow. A financial-data provider may need to compete not only for human subscriptions, but for inclusion inside machine-driven workflows.

The result is a new form of financial distribution in which the decision layer becomes increasingly machine-mediated.

This is one of the reasons Agentic Finance could prove more consequential than simply adding AI features to existing apps.

The Agent can sit above several financial applications rather than belonging completely to one of them.

If that layer becomes the primary place where decisions are made, then value shifts toward whoever controls the objective, the context and the judgment behind those decisions.

Financial Optimization Is Not Comparison Shopping

There is an obvious temptation to describe this future as a world in which Agents automatically move money to whichever product offers the highest return.

That interpretation is too simplistic.

Finance is not ordinary comparison shopping because the object being compared is risk-adjusted capital allocation.

A higher savings rate may come with weaker liquidity. A lending protocol offering higher yield may introduce more credit or smart-contract risk. A trading venue with lower nominal fees may have worse market depth. A portfolio rebalance may improve diversification while creating a large tax liability.

The correct financial decision depends on the relationship between return, risk, liquidity, time horizon, constraints and user objectives.

This means a financial Agent cannot be built around optimization alone.

It requires judgment.

A system that simply maximizes yield may systematically choose the wrong outcomes. A system that only minimizes volatility may fail to meet long-term return objectives. A system that reduces transaction costs without understanding portfolio construction may optimize execution while degrading the strategy.

The Agent therefore needs an objective function that is financially coherent.

This is where the problem moves beyond automation.

Automation asks whether a system can perform an action.

Financial intelligence asks whether the action should be performed in the first place.

As execution becomes easier, this second question becomes more important, not less.

Execution Is Becoming Infrastructure

The infrastructure required to make financial Agents operational is improving quickly.

Brokerages increasingly expose structured APIs. Model Context Protocol and similar standards make external tools easier for AI systems to access. Wallet infrastructure allows more granular permissions. Onchain markets provide machine-readable financial primitives. Account data is becoming easier to integrate into AI workflows.

These developments matter because they reduce the technical distance between an AI recommendation and a financial action.

But they also make execution less differentiated.

If several Agents can eventually access the same market data, the same exchange and the same execution API, then access alone cannot explain why one Agent should outperform another.

The scarce layer moves upward.

The relevant questions become more familiar to professional investors than to software engineers.

Which signals deserve attention?

Which sources are reliable?

How should evidence be weighted?

What invalidates the original thesis?

When should risk increase or decrease?

How should capital be sized?

Under what conditions should the system do nothing?

These are questions of financial judgment.

They cannot be solved simply by giving a model more tools.

Financial Agents Need a Source of Judgment

This is particularly important because there are two very different visions of financial AI.

One assumes that a sufficiently capable general-purpose model will eventually generate superior investment judgment on its own.

The other treats AI primarily as a mechanism for structuring, extending and scaling financial judgment that already exists.

At Questflow, we are much more interested in the second model.

Experienced investors and portfolio managers already operate with differentiated frameworks. They have learned through repeated market exposure which information matters, which signals are misleading, how conditions interact, and when a strategy should not be used.

Much of that knowledge is not written down in a form that can be easily distributed.

It exists in research routines, portfolio decisions, conversations, risk rules and accumulated experience.

AI creates a way to turn more of that tacit judgment into a structured operating system.

An Agent can monitor the information sources the investor cares about. It can identify situations that match the framework. It can explain why conditions have changed. It can apply risk limits consistently. It can preserve the methodology when the investor is not actively watching the market.

The Agent does not need to pretend that the model itself is the source of alpha.

The investor remains the source of judgment.

AI makes the judgment more scalable.

This distinction becomes particularly important as Agents gain greater authority over capital. The more easily a system can act, the more important it becomes to know what framework governs that action.

Permission Must Encode Investment Logic

The next important design problem is therefore not simply whether an Agent has permission to trade.

Financial permissions need to express the structure of the mandate itself.

A user might allow an Agent to move cash among pre-approved institutions but not into risky assets. Another might allow equity trading while limiting individual positions to a maximum percentage of portfolio value. A derivatives strategy may be permitted only for hedging. A portfolio manager may allow an Agent to execute an existing strategy automatically while reserving any change in strategy for explicit approval.

These are not generic software permissions.

They are financial constraints.

Capital allocation, leverage, position sizing, asset eligibility, drawdown thresholds, excluded markets and stop conditions collectively define what the Agent is allowed to do.

This is why unlimited autonomy is a poor model for Agentic Finance.

The useful form of autonomy is scoped.

The Agent acts independently within a mandate that remains transparent and revocable.

That structure mirrors how professional finance already works. Portfolio managers do not generally operate with unlimited freedom. They work within mandates, risk budgets and investment guidelines.

Agentic systems will likely need the same architecture.

The Margin Compression Could Extend Well Beyond Deposits

Once viewed through this framework, the potential effect on finance becomes much broader than the banking example.

Any business model that benefits from customers failing to continuously optimize becomes more exposed.

Brokerage cash is one case. Many investors hold uninvested balances without regularly checking whether those balances could earn more elsewhere.

Credit is another. Refinancing decisions are infrequent because they involve time, paperwork and uncertainty. An Agent can monitor when refinancing becomes economically rational rather than relying on the borrower to notice.

Execution quality is another. Retail investors rarely compare venue-level execution economics manually. Software can.

Asset-management fees could also become easier to challenge if an Agent can continuously compare portfolios against lower-cost alternatives with similar exposures.

Foreign-exchange spreads, insurance products and other categories may eventually face similar pressure.

The effect will not be uniform because many financial margins compensate providers for real services, balance-sheet risk, compliance or capital requirements.

But margins sustained primarily by complexity and inattention will become less defensible.

An Agent makes hidden friction measurable.

Once friction is measurable, it becomes easier to compete away.

From Financial Apps to Financial Mandates

The most interesting long-term implication may therefore be a change in what users actually choose.

Today, consumers mostly choose financial applications.

They choose a bank, brokerage, exchange, wallet or wealth manager.

In a more agentic financial system, they may increasingly choose mandates.

A user could define a requirement such as maintaining six months of liquidity while optimizing the remainder of their cash.

Another could authorize an Agent to follow a portfolio manager’s investment framework while limiting maximum drawdown and leverage.

Another could ask for globally diversified equity exposure with explicit rules for reducing risk during certain macroeconomic conditions.

The interface shifts from choosing every individual action to defining the financial objective.

The underlying applications remain important, but they become tools used by the Agent to satisfy the mandate.

This reverses the conventional super-app model.

A super app tries to bring every financial function into one interface.

An Agent does not necessarily need every function inside one application.

It can coordinate multiple financial services from above.

That means the strategic value may increasingly sit not in aggregating every product into one interface, but in building the intelligence capable of selecting among those products coherently.

Agentic Finance Changes the Meaning of Financial Intelligence

This brings the discussion back to a more fundamental issue.

If the future Agent is expected to monitor, compare and eventually act continuously, then financial intelligence can no longer be defined primarily as the ability to answer finance questions.

It requires a broader set of capabilities.

The Agent needs to understand the user’s objective. It needs to interpret market information. It needs to search across alternatives. It needs to evaluate trade-offs. It needs to reason about risk. It needs to understand when conditions have changed and when a previous decision is no longer valid. It needs to operate within clear permissions.

Most importantly, it needs a coherent source of judgment.

The strongest financial Agents will not necessarily be those that take the most actions or process the most information.

They will be the systems that know which information matters and when action is justified.

This is the layer we believe will become increasingly valuable as execution infrastructure matures.

The first generation of financial AI reduced the cost of answering questions.

The next generation will reduce the cost of maintaining financial attention.

That sounds like a modest change, but economically it is profound.

Attention is one of the hidden constraints that has shaped consumer finance for decades.

Once it becomes cheap, persistent and programmable, many assumptions about financial distribution begin to change.

Capital becomes more mobile.

Products become more comparable.

Institutions must compete more continuously.

The relationship between the customer and the financial provider becomes less sticky.

And the intelligence deciding where capital should move becomes more important than the interface through which the transaction is executed.

The End of Financial Inertia

The financial system will not transition to this model immediately.

Trust remains a major constraint. Regulation will limit how much authority can be delegated. Users will be cautious about allowing software to move money. Legacy systems remain fragmented, and many financial products are still poorly suited to machine-driven comparison.

But the direction is becoming easier to see.

The cost of monitoring money is falling.

The cost of comparing financial alternatives is falling.

The technical cost of executing financial decisions is falling.

As each of these costs declines, capital becomes harder to capture simply through inertia.

That does not imply the end of banks, brokers or asset managers. It implies that these institutions will increasingly need to compete on the actual economic value of what they provide rather than relying as heavily on friction.

The first major shift in fintech democratized access.

Agentic Finance may democratize something more subtle: continuous financial attention.

Professional institutions have always had systems and people watching capital.

Most individuals have not.

If AI makes that capability broadly available, the difference between institutional and retail finance begins to narrow in a new way.

Every dollar does not need to move continuously.

But every dollar can increasingly be watched continuously.

And once capital can be watched continuously, the old economics of financial inertia begin to weaken.

The strategic competition then moves upward—from owning the account, to owning the interface, to owning the Agent, and ultimately to owning the judgment that governs the Agent.

That final layer may become one of the most valuable positions in the financial stack.

Because in a world where execution is abundant and capital is increasingly mobile, the hardest problem is no longer how to move money.

It is deciding where money should go.

Financial AI is changing faster than the way we measure it.

For years, evaluating financial language models was relatively straightforward. A model could be asked to retrieve a figure from a filing, calculate a ratio, summarize an earnings report, explain a financial concept, or produce an investment memo. These tasks remain useful because they test whether a model can read financial information, reason over documents, and produce outputs that resemble the work of an analyst.

But they increasingly describe only part of what financial AI is becoming.

The current generation of financial systems is moving beyond isolated question answering. Models are gaining access to professional data sources, research tools, financial workflows, portfolio information and, increasingly, trading infrastructure.

Kimi’s recently launched financial industry solution is a clear example of this shift. It combines more than ten professional data sources with dedicated financial Skills covering areas such as financial modeling, research reports, portfolio reviews and earnings analysis. Its stated objective is not simply to answer financial questions, but to connect data retrieval, analysis, modeling and professional deliverables into an end-to-end workflow.

A similar transition is taking place closer to execution.

Interactive Brokers now allows users to connect AI systems including ChatGPT, Claude and Grok to authorized brokerage-account data through MCP. These systems can analyze holdings, cash balances, P&L, historical transactions, options data and risk exposures, and can generate trade instructions for the user to review before submission.

These developments suggest that the unit of financial AI is changing.

It is moving from:

answering a financial question

toward:

performing a financial task.

That distinction has significant implications for how financial AI should be evaluated.

Financial Knowledge Is Not the Same as Financial Work

A model may understand what revenue growth means.

It may know the formula for CAGR.

It may be able to explain the difference between EBITDA and free cash flow.

None of these capabilities necessarily demonstrates that the system can perform a real financial workflow reliably.

Consider a relatively ordinary market-screening request:

Find U.S. equities with a market capitalization above $2 billion, average 20-day dollar volume above $50 million, current relative volume above 1.5, and a close above the previous 30-day high. Then rank the qualifying companies by relative volume and return the top 20.

There is nothing conceptually exotic about this task.

But completing it correctly requires a model or Agent to perform several different operations.

It needs to identify the relevant universe, obtain the appropriate data, understand how each metric is defined, calculate derived values, apply multiple filters, identify a technical event, sort the resulting set and return the correct ranking.

A model can fail at any stage.

It can retrieve stale data.

It can calculate the denominator incorrectly.

It can include the current session in a historical lookback that should exclude it.

It can identify the correct securities but rank them incorrectly.

It can omit qualifying results.

It can return securities that never satisfied the original criteria.

The final answer may look entirely plausible while still being materially wrong.

The same problem becomes more pronounced in crypto and onchain markets.

A user might ask for tokens satisfying a combination of liquidity, market capitalization, netflow, active-wallet growth and transaction-volume conditions over a particular time period.

Or an investor may ask an Agent to identify wallets with a defined profitability profile, specific transaction behavior and exposure to a particular category of assets.

These tasks are not primarily tests of financial vocabulary.

They test whether a system can convert a financial objective into a sequence of correct operations over real data.

This is why the distinction between financial knowledge and financial work matters.

The former asks whether a model understands finance.

The latter asks whether it can produce a correct, verifiable financial result.

Financial Benchmarks Have Already Begun to Move in This Direction

The benchmark landscape is not static.

FinanceBench and earlier financial QA datasets established an important foundation by testing models against financial documents, numerical reasoning and fixed reference answers.

More recent benchmarks have expanded the scope considerably.

Vals’ Finance Agent benchmark evaluates systems that can use web search, SEC EDGAR, document parsing, stored information and historical market prices to answer complex financial questions. The model is no longer expected to rely entirely on information contained in the prompt. It must decide how to use tools to retrieve and analyze the required information.

Rogo’s Big Finance Benchmark goes further into professional workflow evaluation. Its 928 questions cover areas including capital structure, valuation, private capital, earnings analysis and forecasting. Rather than grading only the final answer, Rogo breaks complex work into detailed criteria covering retrieval, definitions and calculations, making it possible to identify where a model’s process failed.

This is an important evolution.

A model that retrieved the right filing but made a calculation error should not necessarily be treated as equivalent to a model that failed to identify the correct source entirely.

Likewise, two models with similar aggregate scores can have very different capability profiles across financial workflows. Rogo’s results illustrate this clearly: different frontier models lead in different classes of financial work rather than one model dominating every category.

These benchmarks increasingly measure how AI performs work rather than simply how much it knows.

But the next transition in financial AI introduces another problem.

The market itself is not static.

Financial Agents Operate in a Changing World

Most benchmark design benefits from fixed questions and fixed answers.

That is relatively easy when evaluating document analysis.

A company’s reported revenue for a completed fiscal year does not change every hour.

A historical filing can be frozen.

A calculation performed using defined inputs can be reproduced.

Markets are different.

Prices change.

Volumes change.

Liquidity changes.

Wallet balances change.

Positions change.

Signals appear and disappear.

Rankings can change within minutes.

This becomes a fundamental problem once financial AI is expected to perform market screening, signal detection, cross-asset analysis or portfolio-aware workflows.

Suppose two models receive the same request:

Find the 20 U.S. equities with the strongest relative volume among stocks currently breaking above their previous 30-day high.

Model A executes the task at 10:00 a.m.

Model B executes it at 11:30 a.m.

Even if both systems perform perfectly, their answers may differ because the underlying market has changed.

A benchmark cannot meaningfully compare those outputs unless both models are evaluated against the same financial reality.

This means future financial-agent benchmarks will increasingly need a concept that traditional static QA benchmarks rarely require:

a reproducible market state.

For dynamic financial tasks, the benchmark must specify not only the question but also the moment in the market to which the question refers.

That may require recording the timestamp, data cutoff, exchange session, market-data version and, for onchain tasks, the relevant block or slot.

All evaluated models should then operate against the same frozen state.

The principle is simple:

Same market. Same tools. Same tasks. Different models.

Without this control, performance differences can easily become differences in data rather than differences in intelligence.

Dynamic Markets Require Reproducible Ground Truth

Once financial benchmarks move into dynamic markets, another issue becomes more important: how the correct answer is created.

In many language-model benchmarks, another model is used as the evaluator.

This approach can be useful when the task is genuinely qualitative.

But it should not become the default source of truth for financial questions that can be verified deterministically.

If the task asks for a revenue CAGR, the benchmark should not need an LLM to decide whether the answer is correct.

The required financial figures can be obtained from an authoritative data source and the calculation performed in code.

If the task asks for securities satisfying a set of screening conditions, a machine rule can be executed against a frozen dataset to produce the correct set.

If the task asks for a ranking, the ordering can be calculated from the underlying metric.

In such cases, the benchmark can construct a reproducible Gold Answer independently of the model being tested.

A robust architecture might look like:

Source Data → Standardized Metrics → Frozen Snapshot → Code → Gold Answer

This matters for more than technical elegance.

Financial evaluation needs provenance.

If a benchmark says that a model is wrong, it should be possible to answer several questions:

Where did the underlying data come from?

What definition of the metric was used?

What market state applied?

What calculation produced the reference answer?

Can the same answer be reproduced later?

Which part of the model’s output was incorrect?

For financial applications, these questions are essential.

Accuracy without traceability is difficult to trust.

Not Every Financial Task Has a Deterministic Answer

There is, however, an important limit to this approach.

Some forms of financial judgment cannot be reduced to a formula.

Chart interpretation is one example.

Experienced discretionary investors often recognize patterns that are difficult to describe through one technical indicator. They may evaluate price structure, volume, consolidation, breakout quality, prior volatility and broader market context simultaneously.

Two charts can satisfy the same simple numerical conditions while only one represents the pattern an experienced trader would actually consider actionable.

The same issue exists in other forms of expert financial analysis.

A model may retrieve all the correct information but still fail to interpret its significance in the way an experienced investor would.

This is where expert labels remain important.

A future financial benchmark will likely require several different forms of ground truth rather than one universal grading mechanism.

For deterministic financial calculations, code should provide the answer.

For screening and ranking tasks, machine-executed rules should provide the answer.

For tasks involving expert visual or investment judgment, carefully constructed human labels may remain the appropriate reference.

The general principle should be:

Use the most objective source of truth available for each task.

AI can assist with task generation, rewriting, quality assurance and error analysis.

It should not automatically become the ultimate authority when a stronger source of truth exists.

From Financial Analyst Benchmarks to Financial Agent Benchmarks

As financial AI evolves, the capability map being tested will need to broaden with it.

The relevant question will no longer be whether a model can answer finance questions in the abstract.

A useful benchmark will increasingly need to evaluate whether an Agent can reliably perform several different classes of work.

Retrieval remains fundamental. A system must locate the correct filing, market datapoint or transaction record rather than inventing one.

Calculation tests whether the model can transform raw facts into derived financial metrics using the correct definitions and periods.

Screening measures whether an Agent can search a large universe and return the complete set of objects satisfying several simultaneous constraints.

Ranking adds another layer. Finding the correct candidates is insufficient if the system cannot order them according to the requested financial criterion.

Detection evaluates whether an Agent can identify market events or changes such as breakouts, unusual volume, capital flows or portfolio-risk changes.

Interpretation tests whether the system can explain the financial meaning of information rather than merely extracting it.

Multimodal understanding becomes increasingly important as Agents encounter charts, tables, filings, dashboards and other visual information.

Tool use tests whether the Agent knows when it should retrieve data, search a filing, query a market database or perform a calculation instead of relying on its internal model knowledge.

State awareness matters because financial work happens over time. A useful Agent should understand that its answer depends on a specific market state and update when that state changes.

Eventually, another capability will become increasingly important:

action.

As AI systems gain access to portfolios and financial accounts, the boundary between analysis and execution is becoming thinner.

Interactive Brokers’ current AI integrations are deliberately designed to keep execution under user control. Connected AI systems can inspect authorized portfolio information and generate instructions, but the investor reviews and submits each order.

That human-in-the-loop model is likely to remain important, particularly as financial agents mature.

But even before autonomous execution becomes a benchmark category, the industry needs to understand whether the analysis that precedes an action is reliable.

An Agent should not be considered capable merely because it can call an execution tool.

The more important question is whether the financial judgment leading to the action can withstand verification.

The Environment Matters as Much as the Model

There is another consequence of this shift.

Model evaluation increasingly cannot be separated from the environment in which the model operates.

A frontier model without access to professional market data may perform worse than a smaller model equipped with well-designed financial tools.

A model given poor tool descriptions may fail even if its underlying reasoning capability is strong.

One Agent may have access to real-time pricing while another relies on public web search.

Comparing them directly would measure the environment as much as the model.

This is why the concept of a common financial Agent harness becomes increasingly important.

When the objective is to compare base models, the surrounding conditions should be controlled.

Each model should receive the same available data, the same tools, the same financial Skills, the same task, the same market snapshot and comparable runtime constraints.

Only then can the resulting difference reasonably be attributed to how effectively the model operates within that financial environment.

This is also why leaderboard scores should be treated as contextual measurements rather than permanent properties of a model.

Models change.

Providers update them.

Data environments change.

Tools improve.

A score should therefore mean something closer to:

Model X achieved this result in Benchmark Batch Y, using Environment Z, against Market Snapshot T.

That description may be less elegant than a single permanent leaderboard number.

It is also substantially more accurate.

Financial AI Is Becoming Portfolio-Aware

The recent expansion of financial AI makes this problem increasingly urgent.

Kimi’s new financial stack reflects the movement from general-purpose intelligence toward domain-specific workflow systems. The model is paired with authoritative data sources, Skills, financial modeling capabilities and institutional deliverables.

Interactive Brokers represents a different part of the same transition. ChatGPT, Claude, Grok and other MCP-compatible systems can now operate with authorized context from a real brokerage account, including positions, cash balances, historical transactions and portfolio risks.

The convergence is clear.

Financial AI is moving through several stages:

Model → Tools → Professional Data → Portfolio Context → Financial Agent

The benchmark must follow the same trajectory.

If evaluation remains centered on static document questions while financial products are moving toward continuous market analysis and portfolio-aware workflows, benchmark performance will become increasingly disconnected from actual user experience.

The question is no longer merely whether a model performs well on financial text.

It is whether it performs reliably in a financial environment.

A Benchmark Is Also a Definition of the Product

There is a broader reason this matters.

Benchmarks do more than measure systems after they are built.

They influence what systems are optimized to become.

If a benchmark primarily rewards financial-document extraction, model developers will optimize for financial-document extraction.

If it evaluates multi-step professional workflows, models will improve at those workflows.

If it eventually includes live-market screening, onchain analysis, chart interpretation, risk reasoning and portfolio-aware tool use, the resulting systems will evolve in a different direction.

In that sense, the benchmark becomes a specification for what the industry means by a financial Agent.

This is one of the questions we have been exploring at Questflow.

As we think about financial Agents that can eventually help investors research markets, interpret signals, understand expert frameworks and operate across multiple financial environments, we also need a more rigorous way to understand what these systems can actually do.

We are therefore increasingly interested in benchmark architectures centered on real financial tasks, reproducible market states, machine-verifiable outcomes and expert judgment where deterministic answers are not available.

The objective is not simply to create another collection of financial questions.

It is to understand whether an AI system can reliably transform a financial request into a correct and verifiable piece of work.

That distinction becomes increasingly important as the consequences of AI outputs move closer to real capital.

Measuring Financial Intelligence in the Agentic Era

Financial AI is entering a different phase.

The first generation demonstrated that language models could understand financial documents.

The second showed that Agents could use tools to perform increasingly realistic analyst workflows.

The emerging generation is beginning to operate with professional data, continuously changing markets, real portfolio context and increasingly direct access to financial infrastructure.

That progression requires a corresponding change in evaluation.

Static questions will remain useful.

Financial documents will remain important.

Traditional analyst work will remain a core capability.

But they will no longer be sufficient on their own.

The next generation of financial benchmarks will need to measure how models operate in a changing environment.

They will need to evaluate complete sets rather than convenient examples.

They will need to test ranking as well as retrieval.

They will need to distinguish between data errors, calculation errors, interpretation errors and tool-use failures.

They will need market snapshots for dynamic tasks.

They will need reproducible Gold Answers wherever possible.

They will need expert labels where judgment cannot be reduced to code.

And they will need to make performance sufficiently transparent that a user can understand not only whether an Agent failed, but how it failed.

This is particularly important in finance.

A fluent answer can be wrong.

A convincing investment narrative can be constructed from incorrect inputs.

A model can identify nineteen correct securities and quietly omit the twentieth.

A portfolio recommendation can appear reasonable while relying on stale holdings.

As financial AI gets closer to decision-making and execution, surface-level plausibility matters less.

Verifiability matters more.

The most important financial benchmark of the next generation may therefore not ask:

How well does this model know finance?

It may ask something more demanding:

Given the same financial world, the same tools and the same task, can this Agent consistently produce the correct work?

That is a much higher standard.

It is also increasingly the standard financial AI will need to meet.

Financial AI is changing faster than the way we measure it.

For years, evaluating financial language models was relatively straightforward. A model could be asked to retrieve a figure from a filing, calculate a ratio, summarize an earnings report, explain a financial concept, or produce an investment memo. These tasks remain useful because they test whether a model can read financial information, reason over documents, and produce outputs that resemble the work of an analyst.

But they increasingly describe only part of what financial AI is becoming.

The current generation of financial systems is moving beyond isolated question answering. Models are gaining access to professional data sources, research tools, financial workflows, portfolio information and, increasingly, trading infrastructure.

Kimi’s recently launched financial industry solution is a clear example of this shift. It combines more than ten professional data sources with dedicated financial Skills covering areas such as financial modeling, research reports, portfolio reviews and earnings analysis. Its stated objective is not simply to answer financial questions, but to connect data retrieval, analysis, modeling and professional deliverables into an end-to-end workflow.

A similar transition is taking place closer to execution.

Interactive Brokers now allows users to connect AI systems including ChatGPT, Claude and Grok to authorized brokerage-account data through MCP. These systems can analyze holdings, cash balances, P&L, historical transactions, options data and risk exposures, and can generate trade instructions for the user to review before submission.

These developments suggest that the unit of financial AI is changing.

It is moving from:

answering a financial question

toward:

performing a financial task.

That distinction has significant implications for how financial AI should be evaluated.

Financial Knowledge Is Not the Same as Financial Work

A model may understand what revenue growth means.

It may know the formula for CAGR.

It may be able to explain the difference between EBITDA and free cash flow.

None of these capabilities necessarily demonstrates that the system can perform a real financial workflow reliably.

Consider a relatively ordinary market-screening request:

Find U.S. equities with a market capitalization above $2 billion, average 20-day dollar volume above $50 million, current relative volume above 1.5, and a close above the previous 30-day high. Then rank the qualifying companies by relative volume and return the top 20.

There is nothing conceptually exotic about this task.

But completing it correctly requires a model or Agent to perform several different operations.

It needs to identify the relevant universe, obtain the appropriate data, understand how each metric is defined, calculate derived values, apply multiple filters, identify a technical event, sort the resulting set and return the correct ranking.

A model can fail at any stage.

It can retrieve stale data.

It can calculate the denominator incorrectly.

It can include the current session in a historical lookback that should exclude it.

It can identify the correct securities but rank them incorrectly.

It can omit qualifying results.

It can return securities that never satisfied the original criteria.

The final answer may look entirely plausible while still being materially wrong.

The same problem becomes more pronounced in crypto and onchain markets.

A user might ask for tokens satisfying a combination of liquidity, market capitalization, netflow, active-wallet growth and transaction-volume conditions over a particular time period.

Or an investor may ask an Agent to identify wallets with a defined profitability profile, specific transaction behavior and exposure to a particular category of assets.

These tasks are not primarily tests of financial vocabulary.

They test whether a system can convert a financial objective into a sequence of correct operations over real data.

This is why the distinction between financial knowledge and financial work matters.

The former asks whether a model understands finance.

The latter asks whether it can produce a correct, verifiable financial result.

Financial Benchmarks Have Already Begun to Move in This Direction

The benchmark landscape is not static.

FinanceBench and earlier financial QA datasets established an important foundation by testing models against financial documents, numerical reasoning and fixed reference answers.

More recent benchmarks have expanded the scope considerably.

Vals’ Finance Agent benchmark evaluates systems that can use web search, SEC EDGAR, document parsing, stored information and historical market prices to answer complex financial questions. The model is no longer expected to rely entirely on information contained in the prompt. It must decide how to use tools to retrieve and analyze the required information.

Rogo’s Big Finance Benchmark goes further into professional workflow evaluation. Its 928 questions cover areas including capital structure, valuation, private capital, earnings analysis and forecasting. Rather than grading only the final answer, Rogo breaks complex work into detailed criteria covering retrieval, definitions and calculations, making it possible to identify where a model’s process failed.

This is an important evolution.

A model that retrieved the right filing but made a calculation error should not necessarily be treated as equivalent to a model that failed to identify the correct source entirely.

Likewise, two models with similar aggregate scores can have very different capability profiles across financial workflows. Rogo’s results illustrate this clearly: different frontier models lead in different classes of financial work rather than one model dominating every category.

These benchmarks increasingly measure how AI performs work rather than simply how much it knows.

But the next transition in financial AI introduces another problem.

The market itself is not static.

Financial Agents Operate in a Changing World

Most benchmark design benefits from fixed questions and fixed answers.

That is relatively easy when evaluating document analysis.

A company’s reported revenue for a completed fiscal year does not change every hour.

A historical filing can be frozen.

A calculation performed using defined inputs can be reproduced.

Markets are different.

Prices change.

Volumes change.

Liquidity changes.

Wallet balances change.

Positions change.

Signals appear and disappear.

Rankings can change within minutes.

This becomes a fundamental problem once financial AI is expected to perform market screening, signal detection, cross-asset analysis or portfolio-aware workflows.

Suppose two models receive the same request:

Find the 20 U.S. equities with the strongest relative volume among stocks currently breaking above their previous 30-day high.

Model A executes the task at 10:00 a.m.

Model B executes it at 11:30 a.m.

Even if both systems perform perfectly, their answers may differ because the underlying market has changed.

A benchmark cannot meaningfully compare those outputs unless both models are evaluated against the same financial reality.

This means future financial-agent benchmarks will increasingly need a concept that traditional static QA benchmarks rarely require:

a reproducible market state.

For dynamic financial tasks, the benchmark must specify not only the question but also the moment in the market to which the question refers.

That may require recording the timestamp, data cutoff, exchange session, market-data version and, for onchain tasks, the relevant block or slot.

All evaluated models should then operate against the same frozen state.

The principle is simple:

Same market. Same tools. Same tasks. Different models.

Without this control, performance differences can easily become differences in data rather than differences in intelligence.

Dynamic Markets Require Reproducible Ground Truth

Once financial benchmarks move into dynamic markets, another issue becomes more important: how the correct answer is created.

In many language-model benchmarks, another model is used as the evaluator.

This approach can be useful when the task is genuinely qualitative.

But it should not become the default source of truth for financial questions that can be verified deterministically.

If the task asks for a revenue CAGR, the benchmark should not need an LLM to decide whether the answer is correct.

The required financial figures can be obtained from an authoritative data source and the calculation performed in code.

If the task asks for securities satisfying a set of screening conditions, a machine rule can be executed against a frozen dataset to produce the correct set.

If the task asks for a ranking, the ordering can be calculated from the underlying metric.

In such cases, the benchmark can construct a reproducible Gold Answer independently of the model being tested.

A robust architecture might look like:

Source Data → Standardized Metrics → Frozen Snapshot → Code → Gold Answer

This matters for more than technical elegance.

Financial evaluation needs provenance.

If a benchmark says that a model is wrong, it should be possible to answer several questions:

Where did the underlying data come from?

What definition of the metric was used?

What market state applied?

What calculation produced the reference answer?

Can the same answer be reproduced later?

Which part of the model’s output was incorrect?

For financial applications, these questions are essential.

Accuracy without traceability is difficult to trust.

Not Every Financial Task Has a Deterministic Answer

There is, however, an important limit to this approach.

Some forms of financial judgment cannot be reduced to a formula.

Chart interpretation is one example.

Experienced discretionary investors often recognize patterns that are difficult to describe through one technical indicator. They may evaluate price structure, volume, consolidation, breakout quality, prior volatility and broader market context simultaneously.

Two charts can satisfy the same simple numerical conditions while only one represents the pattern an experienced trader would actually consider actionable.

The same issue exists in other forms of expert financial analysis.

A model may retrieve all the correct information but still fail to interpret its significance in the way an experienced investor would.

This is where expert labels remain important.

A future financial benchmark will likely require several different forms of ground truth rather than one universal grading mechanism.

For deterministic financial calculations, code should provide the answer.

For screening and ranking tasks, machine-executed rules should provide the answer.

For tasks involving expert visual or investment judgment, carefully constructed human labels may remain the appropriate reference.

The general principle should be:

Use the most objective source of truth available for each task.

AI can assist with task generation, rewriting, quality assurance and error analysis.

It should not automatically become the ultimate authority when a stronger source of truth exists.

From Financial Analyst Benchmarks to Financial Agent Benchmarks

As financial AI evolves, the capability map being tested will need to broaden with it.

The relevant question will no longer be whether a model can answer finance questions in the abstract.

A useful benchmark will increasingly need to evaluate whether an Agent can reliably perform several different classes of work.

Retrieval remains fundamental. A system must locate the correct filing, market datapoint or transaction record rather than inventing one.

Calculation tests whether the model can transform raw facts into derived financial metrics using the correct definitions and periods.

Screening measures whether an Agent can search a large universe and return the complete set of objects satisfying several simultaneous constraints.

Ranking adds another layer. Finding the correct candidates is insufficient if the system cannot order them according to the requested financial criterion.

Detection evaluates whether an Agent can identify market events or changes such as breakouts, unusual volume, capital flows or portfolio-risk changes.

Interpretation tests whether the system can explain the financial meaning of information rather than merely extracting it.

Multimodal understanding becomes increasingly important as Agents encounter charts, tables, filings, dashboards and other visual information.

Tool use tests whether the Agent knows when it should retrieve data, search a filing, query a market database or perform a calculation instead of relying on its internal model knowledge.

State awareness matters because financial work happens over time. A useful Agent should understand that its answer depends on a specific market state and update when that state changes.

Eventually, another capability will become increasingly important:

action.

As AI systems gain access to portfolios and financial accounts, the boundary between analysis and execution is becoming thinner.

Interactive Brokers’ current AI integrations are deliberately designed to keep execution under user control. Connected AI systems can inspect authorized portfolio information and generate instructions, but the investor reviews and submits each order.

That human-in-the-loop model is likely to remain important, particularly as financial agents mature.

But even before autonomous execution becomes a benchmark category, the industry needs to understand whether the analysis that precedes an action is reliable.

An Agent should not be considered capable merely because it can call an execution tool.

The more important question is whether the financial judgment leading to the action can withstand verification.

The Environment Matters as Much as the Model

There is another consequence of this shift.

Model evaluation increasingly cannot be separated from the environment in which the model operates.

A frontier model without access to professional market data may perform worse than a smaller model equipped with well-designed financial tools.

A model given poor tool descriptions may fail even if its underlying reasoning capability is strong.

One Agent may have access to real-time pricing while another relies on public web search.

Comparing them directly would measure the environment as much as the model.

This is why the concept of a common financial Agent harness becomes increasingly important.

When the objective is to compare base models, the surrounding conditions should be controlled.

Each model should receive the same available data, the same tools, the same financial Skills, the same task, the same market snapshot and comparable runtime constraints.

Only then can the resulting difference reasonably be attributed to how effectively the model operates within that financial environment.

This is also why leaderboard scores should be treated as contextual measurements rather than permanent properties of a model.

Models change.

Providers update them.

Data environments change.

Tools improve.

A score should therefore mean something closer to:

Model X achieved this result in Benchmark Batch Y, using Environment Z, against Market Snapshot T.

That description may be less elegant than a single permanent leaderboard number.

It is also substantially more accurate.

Financial AI Is Becoming Portfolio-Aware

The recent expansion of financial AI makes this problem increasingly urgent.

Kimi’s new financial stack reflects the movement from general-purpose intelligence toward domain-specific workflow systems. The model is paired with authoritative data sources, Skills, financial modeling capabilities and institutional deliverables.

Interactive Brokers represents a different part of the same transition. ChatGPT, Claude, Grok and other MCP-compatible systems can now operate with authorized context from a real brokerage account, including positions, cash balances, historical transactions and portfolio risks.

The convergence is clear.

Financial AI is moving through several stages:

Model → Tools → Professional Data → Portfolio Context → Financial Agent

The benchmark must follow the same trajectory.

If evaluation remains centered on static document questions while financial products are moving toward continuous market analysis and portfolio-aware workflows, benchmark performance will become increasingly disconnected from actual user experience.

The question is no longer merely whether a model performs well on financial text.

It is whether it performs reliably in a financial environment.

A Benchmark Is Also a Definition of the Product

There is a broader reason this matters.

Benchmarks do more than measure systems after they are built.

They influence what systems are optimized to become.

If a benchmark primarily rewards financial-document extraction, model developers will optimize for financial-document extraction.

If it evaluates multi-step professional workflows, models will improve at those workflows.

If it eventually includes live-market screening, onchain analysis, chart interpretation, risk reasoning and portfolio-aware tool use, the resulting systems will evolve in a different direction.

In that sense, the benchmark becomes a specification for what the industry means by a financial Agent.

This is one of the questions we have been exploring at Questflow.

As we think about financial Agents that can eventually help investors research markets, interpret signals, understand expert frameworks and operate across multiple financial environments, we also need a more rigorous way to understand what these systems can actually do.

We are therefore increasingly interested in benchmark architectures centered on real financial tasks, reproducible market states, machine-verifiable outcomes and expert judgment where deterministic answers are not available.

The objective is not simply to create another collection of financial questions.

It is to understand whether an AI system can reliably transform a financial request into a correct and verifiable piece of work.

That distinction becomes increasingly important as the consequences of AI outputs move closer to real capital.

Measuring Financial Intelligence in the Agentic Era

Financial AI is entering a different phase.

The first generation demonstrated that language models could understand financial documents.

The second showed that Agents could use tools to perform increasingly realistic analyst workflows.

The emerging generation is beginning to operate with professional data, continuously changing markets, real portfolio context and increasingly direct access to financial infrastructure.

That progression requires a corresponding change in evaluation.

Static questions will remain useful.

Financial documents will remain important.

Traditional analyst work will remain a core capability.

But they will no longer be sufficient on their own.

The next generation of financial benchmarks will need to measure how models operate in a changing environment.

They will need to evaluate complete sets rather than convenient examples.

They will need to test ranking as well as retrieval.

They will need to distinguish between data errors, calculation errors, interpretation errors and tool-use failures.

They will need market snapshots for dynamic tasks.

They will need reproducible Gold Answers wherever possible.

They will need expert labels where judgment cannot be reduced to code.

And they will need to make performance sufficiently transparent that a user can understand not only whether an Agent failed, but how it failed.

This is particularly important in finance.

A fluent answer can be wrong.

A convincing investment narrative can be constructed from incorrect inputs.

A model can identify nineteen correct securities and quietly omit the twentieth.

A portfolio recommendation can appear reasonable while relying on stale holdings.

As financial AI gets closer to decision-making and execution, surface-level plausibility matters less.

Verifiability matters more.

The most important financial benchmark of the next generation may therefore not ask:

How well does this model know finance?

It may ask something more demanding:

Given the same financial world, the same tools and the same task, can this Agent consistently produce the correct work?

That is a much higher standard.

It is also increasingly the standard financial AI will need to meet.

For much of crypto’s history, the simplest way to invest in the growth of an ecosystem was to own the infrastructure beneath it.

If applications attracted users, the chain processed more transactions. If transaction demand increased, fees increased. If more economic activity settled on the network, the native asset was expected to capture some portion of that growth. This logic helped define several crypto cycles and supported enormous valuations across Layer 1 networks, Layer 2 systems, interoperability protocols, data infrastructure, and other components of the blockchain stack.

That relationship is becoming less straightforward.

One of the more revealing debates of 2026 began when Bankless co-founder David Hoffman sold the remainder of his ETH position. His argument was not that Ethereum had failed. In fact, he remained explicitly optimistic about Ethereum as a network. His concern was instead about value capture: Ethereum could continue becoming more useful, more secure, and more economically important without that success necessarily producing a proportional revaluation of ETH.

That distinction has become increasingly relevant over the past several months.

While major crypto assets spent much of the year struggling to establish a clear direction, a number of application-oriented assets began performing according to their own economics. Venice’s VVV, Hyperliquid, Pump, EtherFi and other projects increasingly attracted attention not simply because they belonged to a particular ecosystem, but because investors could point to users, revenue, fees, buybacks, or other mechanisms through which economic activity might accrue to the asset itself.

The market may be beginning to ask a different question.

Not simply: Which blockchain will win?

But: Where in the stack will the economic value ultimately accumulate?

The recent rise of Robinhood Chain makes that question much harder—and much more interesting.

From the Infrastructure Thesis to the Application Thesis

Austin Barack, a former CoinFund partner and the founder of Relayer Capital, recently described the change in unusually direct terms during a conversation with Hoffman.

For much of crypto’s history, he argued, execution infrastructure generated more than 95% of the industry’s revenue. Today, applications account for roughly two-thirds, while execution layers represent closer to one-third. His expectation is that this transition continues, potentially leaving applications responsible for more than 90% of crypto revenue over time.

The exact percentages will inevitably fluctuate, but the structural argument deserves attention.

Blockchains have spent years making execution cheaper.

Ethereum’s rollup roadmap is explicitly designed to expand blockspace. Alternative Layer 1s compete on throughput and cost. Layer 2 networks compress transactions further. Data availability becomes cheaper. Wallet infrastructure improves. Liquidity becomes portable.

These developments are excellent for users and applications.

They are less obviously positive for infrastructure margins.

This is a familiar pattern in technology. As infrastructure matures, it becomes standardized, increasingly interchangeable, and progressively cheaper. Value then tends to migrate toward the businesses that own distribution, users, workflows, proprietary information, or differentiated experiences.

Cloud computing did not eliminate infrastructure. It made infrastructure abundant enough for enormous application businesses to be built on top of it.

Crypto may now be entering a similar phase.

The chain remains essential. But being essential does not necessarily mean capturing most of the economics.

VVV, HYPE and the Emergence of Crypto Businesses

The recent interest in assets such as VVV is instructive.

Venice is fundamentally an AI application with a direct consumer relationship. Users purchase subscriptions and inference credits. Its token economics include programmatic mechanisms that direct part of that business activity toward VVV burns. Austin’s investment thesis therefore resembles an analysis of a technology business considerably more than the traditional crypto framework of measuring transactions, TVL, or ecosystem developers.

The same analytical shift appears in discussions around Hyperliquid, Pump and EtherFi.

Instead of asking primarily which ecosystem they belong to, investors increasingly ask questions that would be recognizable in conventional business analysis: How quickly is revenue growing? How durable is the customer base? What are the margins? Is revenue cyclical? How much economic value is returned to token holders? How much must be reinvested to sustain growth?

Austin described these assets as partially decoupling from the broader crypto market precisely because their underlying businesses were growing independently and returning measurable value to token holders.

That does not make these assets equivalent to equities. Token holders frequently lack the legal claims associated with shareholders, and value-accrual policies can change. Those differences remain substantial.

But the analytical direction is important.

Crypto is slowly moving from valuing infrastructure that might eventually host economic activity toward valuing applications that already have economic activity.

Robinhood Chain makes this transition particularly visible.

What Robinhood Chain’s Fee Surge Actually Tells Us

During the first week of September, Robinhood Chain became one of the most discussed networks in crypto.

On September 4, the network generated approximately $6 million in daily chain fees, its highest level to date. Fees across the preceding seven days reached roughly $25 million, compared with around $1.4 million during the prior week. DEX volume more than doubled to approximately $12.4 billion. At various points Robinhood Chain surpassed Ethereum, Base and other major networks in daily chain fees.

The obvious interpretation is that Robinhood Chain is rapidly taking market share.

But the underlying data suggests a more nuanced story.

The increase was overwhelmingly associated with Pons, a token launchpad that became the dominant application on the network. On September 3, Pons itself generated close to $6 million in fees, exceeding the daily fee generation of several much larger crypto applications. Meanwhile, average daily active accounts on Robinhood Chain actually declined during the week in which fees surged.

This distinction matters.

Robinhood Chain did not suddenly acquire seventeen times as many users.

Instead, the economic activity occurring through applications on the chain became dramatically more valuable.

That is almost a perfect illustration of the application thesis.

The chain is necessary to make the activity possible, yet much of the economic differentiation sits one layer above it.

Current DefiLlama data makes the separation even more visible. Robinhood Chain presently generates significant chain fees, but applications running on the network collectively generate substantially larger application fees.

The emerging question is therefore not simply whether Robinhood Chain can generate blockspace demand.

It is who owns the profitable activity creating that demand.

Is Robinhood Chain Eating Base?

The rapid increase in Robinhood Chain activity has naturally produced comparisons with Base.

The argument is understandable.

Both sit within the broader Ethereum scaling ecosystem. Both are associated with major consumer financial platforms. Coinbase built Base around one of crypto’s largest existing distribution channels; Robinhood arrives with a large brokerage user base and an increasingly aggressive strategy around tokenized equities, prediction markets, crypto trading and financial infrastructure.

If Robinhood can bring its users, assets and financial products onto its own chain, why should that activity happen on Base?

There is likely to be some competition.

Tokenized equities are an obvious example. Robinhood has a strong incentive to make Robinhood Chain the native environment for financial products it distributes. Coinbase has precisely the same incentive for Base.

But describing the situation as Robinhood “eating Base” is premature.

Base still processed approximately 10.5 million daily transactions in the latest Growthepie data, up nearly 28% week over week, and held approximately $4.9 billion in stablecoins. Robinhood Chain recorded roughly 9.6 million transactions and around $1 billion in stablecoins. Base’s transaction activity has therefore not collapsed alongside its relative decline in fee rankings.

This exposes a flaw in treating chain fees as a complete measure of ecosystem strength.

A low-fee chain can process enormous economic activity while intentionally capturing relatively little of it.

A high-fee period can also reflect temporary congestion or an unusually profitable speculative application rather than durable network dominance.

Fees measure extraction.

They do not necessarily measure distribution.

And distribution may be the more important strategic asset.

Robinhood and Base Are Not Really Chain Businesses

This leads to a different way of understanding the competition.

Coinbase’s greatest advantage is not Base’s blockspace.

It is Coinbase.

Robinhood Chain’s greatest advantage is not its consensus architecture.

It is Robinhood.

Both companies already have something that most blockchain ecosystems spent billions of dollars trying to manufacture: a direct relationship with financial users.

That relationship changes the economics of launching a blockchain.

A standalone Layer 1 needs to attract developers, applications, liquidity and eventually users. A financial application with millions of existing customers can reverse the process. It begins with distribution and builds infrastructure underneath it.

The chain becomes an extension of the application rather than the application being dependent on the chain.

This inversion may be one of the most important changes taking place in crypto.

The original model was:

Chain → Applications → Users

The emerging model can increasingly look like:

Distribution → Application → Chain

Robinhood does not need Robinhood Chain to become a destination that users consciously choose because of its technical architecture. It needs the chain to make Robinhood’s financial products more programmable, interoperable and efficient.

The user may eventually have no reason to know which execution environment is being used at all.

That is not necessarily bearish for blockchains.

It is bearish for the assumption that the chain must always capture the majority of the value created above it.

The Application Layer May Eventually Become the Agent Layer

We believe this transition becomes even more significant when AI Agents enter the financial stack.

Today’s application thesis assumes that users interact with applications directly.

They open Hyperliquid to trade.

They open Robinhood to invest.

They use a wallet to move assets.

They enter a lending application to borrow.

They visit an AI product to conduct research.

But Agentic Finance begins changing this relationship.

A financial Agent can potentially interact with several of these systems simultaneously. It can monitor markets, analyze information, compare opportunities, access liquidity, manage risk and execute through whichever financial venue best satisfies its mandate.

In that environment, even the application interface begins to become infrastructure.

The scarce layer moves upward again.

Consider what happens if a user no longer chooses between Robinhood Chain, Base, Ethereum, Hyperliquid or a prediction market before making an investment decision.

Instead, the user chooses a financial Agent.

The Agent determines which information matters, identifies an opportunity, selects the appropriate market, chooses the execution route and manages the resulting position within permissions established by the user.

The economic hierarchy then changes from:

Chain → Application → User

toward something closer to:

Infrastructure → Financial Applications → Agent → User

The Agent becomes the primary relationship through which the investor experiences the market.

This has profound implications for value capture.

If blockchains compete for applications today, they may eventually compete for Agent order flow.

If exchanges compete for traders today, they may eventually compete to become the preferred execution venue of financial Agents.

And if applications currently own customer relationships, Agents may begin owning the layer at which financial decisions themselves are made.

Distribution Is Becoming More Important Than Blockspace

This is why we do not view Robinhood Chain’s rise primarily as a threat to Base.

It is evidence of something broader.

Crypto is moving from a competition for blockspace toward a competition for financial distribution.

Robinhood has brokerage distribution.

Coinbase has crypto distribution.

Hyperliquid has trading distribution.

Venice has AI-user distribution.

Pump has speculative-asset distribution.

EtherFi is attempting to build a broader financial relationship through payments, borrowing, yield and tokenized assets.

The strongest products increasingly use blockchain infrastructure rather than asking users to care about blockchain infrastructure.

That is a healthier development for the industry.

It also means that the winners of the next cycle may look less like generalized infrastructure protocols and more like vertically integrated financial applications with clear users, identifiable revenue, and differentiated distribution.

Robinhood Chain may take some activity from Base.

It may also expand the market by introducing users and assets that would otherwise never have entered an onchain environment.

Both outcomes can happen simultaneously.

The more important shift is that neither Robinhood nor Coinbase necessarily needs to maximize chain fees to create enormous economic value. Their primary opportunity exists in the products, transactions, assets and user relationships that their chains enable.

The chain is becoming part of the cost structure of distribution.

It is no longer necessarily the product.

What This Means for Financial Intelligence

At Questflow, this leads us to a further conclusion.

If infrastructure becomes abundant and financial applications become increasingly interoperable, execution itself becomes less differentiated.

An AI Agent will eventually be able to trade on multiple venues, access tokenized assets, interact with lending markets and move capital between different financial environments.

The difficult problem will not be giving the Agent access.

It will be determining what intelligence governs its actions.

This is why we believe the application thesis will eventually evolve into a financial-intelligence thesis.

The most valuable financial Agents will not simply be wrappers around execution APIs. Their differentiation will come from the quality of the judgment they can operationalize: which information they monitor, how they interpret it, how they form and invalidate investment theses, how they size risk, and when they decide not to act.

Our view at Questflow is that this judgment does not need to originate from a generic AI model.

Great investors already possess it.

AI makes it possible to capture that judgment, structure it, monitor markets against it continuously, explain it to users, and eventually execute it across different financial venues.

In such a system, the underlying chain matters technically, but increasingly less from the user’s perspective.

The user’s relationship is with the intelligence.

The Next Crypto Premium

David Hoffman’s decision to sell ETH triggered debate because it challenged one of crypto’s longest-standing assumptions: that growth in a network should naturally produce proportional value for its native asset.

The recent performance of application-oriented tokens, Austin Barack’s application-revenue thesis, and Robinhood Chain’s sudden rise all point toward the same structural question.

Where does value ultimately settle when infrastructure becomes cheaper?

Our view is that there will not be one answer.

Money can capture value because it is money.

Infrastructure can capture value when blockspace or security is genuinely scarce.

Applications can capture value through distribution, revenue, network effects and proprietary user relationships.

And increasingly, financial Agents may capture value by owning the layer where judgment becomes action.

The mistake would be to interpret this transition as the end of blockchains.

It is closer to the opposite.

Blockchains may finally be becoming mature enough that users no longer need to think about them.

The infrastructure disappears into the product.

The product disappears into the Agent.

What remains visible is the outcome the user actually wanted in the first place.

For finance, that outcome was never blockspace.

It was better allocation of capital.

Robinhood Chain’s recent growth therefore matters less because it briefly topped a fee ranking and more because it demonstrates how quickly economic activity can migrate when distribution, applications and programmable financial infrastructure converge.

Base is unlikely to disappear. Robinhood Chain is unlikely to absorb the entire market. Ethereum can continue growing even while a greater share of economics accumulates above it.

The larger change is the hierarchy itself.

The infrastructure era rewarded ownership of the rails.
The application era rewards ownership of the user.
The Agentic era may reward ownership of the judgment.

We believe that final layer will become increasingly important as financial markets move onchain and AI becomes capable of operating across them.

Because when assets can move anywhere and execution can happen anywhere, the most valuable question is no longer where the transaction settles.

It is who—or what—decides that the transaction should happen at all.

Robinhood Chain is easy to interpret as another step in the convergence between traditional finance and crypto. Robinhood has already brought equities to a generation of mobile investors; tokenized stocks now extend that distribution model into blockchain infrastructure. Viewed narrowly, the story is about real-world assets, lower-friction settlement, and the continued migration of financial products onchain.

We believe the more important implication lies elsewhere.

Robinhood Chain is arriving at the same moment that AI is beginning to move beyond financial research and into financial execution. Robinhood itself has made that direction explicit through its work on Agentic Trading, enabling AI systems to interact with market data and trading infrastructure while operating within permissions defined by users.

These two developments—programmable financial assets and increasingly capable financial agents—are not separate trends. Together, they suggest a new architecture for markets in which software is no longer limited to displaying information or routing orders. It can increasingly interpret information, maintain an investment mandate, make decisions, and interact directly with financial infrastructure.

The significance of Robinhood Chain, therefore, may not be that it brings stocks onchain. It may be that it helps create an environment in which financial assets become legible and executable by machines.

That is a much larger transition.

From Market Access to Programmable Finance

The first major wave of consumer fintech was primarily a distribution revolution.

Robinhood reduced the economic and product friction involved in accessing public markets. Crypto exchanges extended that model to globally accessible, continuously operating markets. DeFi went further by replacing parts of the traditional financial intermediary stack with open protocols.

Each generation improved access, but the underlying user model remained largely unchanged. Financial infrastructure presented information and execution tools; the individual investor remained responsible for interpreting the market and deciding what to do.

Blockchain infrastructure introduces a different property: programmability.

Robinhood Stock Tokens, for example, are represented as ERC-20 assets and can therefore interact with the broader logic of smart contracts, wallets, trading venues, lending markets, and other applications. Robinhood Chain is EVM-compatible and incorporates infrastructure such as account abstraction, oracle systems, and cross-chain connectivity.

The relevance of this architecture is not simply that a tokenized asset can move more easily between applications. Programmability means that financial assets can increasingly participate in systems whose actions are initiated and coordinated by software.

Once an asset can be discovered, priced, transferred, traded, collateralized, or combined through machine-readable interfaces, financial activity no longer needs to begin with a person manually navigating a brokerage application.

It can begin with software interpreting a mandate.

This is the point at which tokenization becomes relevant to Agentic Finance.

Financial Markets Are Becoming Machine-Readable

Much of the discussion around tokenization has focused on the benefits to investors: continuous markets, composability, global distribution, improved settlement, and a broader range of financial products available through a wallet.

Those advantages are meaningful, but they still describe the market from a human perspective.

AI Agents require something different. They need structured data, accessible execution, explicit permissions, reliable pricing, and financial primitives that can be invoked programmatically.

Traditional financial infrastructure was not originally designed around these requirements. Market information, brokerage accounts, custody systems, execution venues, and risk controls were built as separate systems, often connected through institution-specific APIs and operational processes.

Onchain finance is gradually compressing many of these functions into a more interoperable environment.

An Agent can potentially read an asset price from an oracle, inspect available liquidity, interact with a smart contract, submit a transaction through a programmable wallet, and operate under predefined authorization constraints. Account abstraction and session-based permissions further make it possible to separate control of an account from unrestricted access to that account.

This does not eliminate the complexity of financial markets. It changes who can interact with that complexity.

The infrastructure becomes increasingly compatible with autonomous software.

Robinhood’s move toward Agentic Trading reinforces this direction. The broader objective is no longer simply to provide an AI assistant that summarizes financial information. It is to allow an Agent to continuously process market data and, within user-defined constraints, connect that analysis to execution.

That distinction is fundamental.

A chatbot produces an answer.

A financial Agent maintains a mandate.

Execution Is Becoming Commoditized. Judgment Is Not.

As AI enters financial markets, there is a tendency to focus on the final step: execution.

Can an Agent trade autonomously? Can it interact with a wallet? Can it submit an order? Can it rebalance a portfolio without requiring approval for every transaction?

These questions matter, but they are unlikely to remain the most difficult part of the system.

Execution infrastructure is becoming increasingly standardized. Exchanges expose APIs. Blockchain networks expose smart contracts. Wallets are becoming programmable. Model Context Protocol and similar interfaces allow AI systems to access external tools in increasingly structured ways.

If this trend continues, the ability of an AI Agent to place a trade will become relatively ordinary.

The harder question is why the trade should exist in the first place.

Financial markets do not reward activity by itself. They reward judgment under uncertainty.

Professional investors develop that judgment through a combination of domain knowledge, experience, proprietary information, pattern recognition, risk management, and repeated exposure to market cycles. Their advantage is rarely reducible to a single signal. It is often a framework for deciding which signals matter under which conditions.

A macro investor may interpret the same inflation print differently depending on positioning, rates expectations, fiscal policy, liquidity, and the market’s prior assumptions. An equity investor may respond to an earnings miss very differently depending on the quality of revenue, unit economics, forward guidance, competitive positioning, and what was already priced into the stock.

The information is observable.

The interpretation is scarce.

This is why we believe the central bottleneck in Agentic Finance will eventually move away from access to models and execution infrastructure and toward the quality of the judgment encoded within the Agent.

An Agent that can trade is not necessarily a financial Agent in the meaningful sense of the term. Without a coherent framework for interpreting markets, it is simply an automated execution system attached to a language model.

Financial Agents Are Not the Same as Trading Bots

Algorithmic trading has existed for decades, so the idea of software executing transactions is not new.

What is changing is the potential scope of the software’s responsibility.

A traditional trading system is usually built around explicit rules or models. Given a defined set of inputs, it calculates a signal and executes according to predetermined logic. Even sophisticated quantitative systems generally operate within a clearly specified strategy architecture.

AI Agents make a different class of system possible.

They can work with unstructured information, combine multiple forms of evidence, maintain contextual memory, use different analytical tools, and reason over conditions that are difficult to express as a single deterministic rule.

This potentially allows an investment framework to be represented at a higher level of abstraction.

Instead of encoding only the instruction to buy when a numerical condition is met, a financial Agent can be designed around an investment thesis: what information should be monitored, what evidence increases conviction, what developments weaken the thesis, which scenarios require further investigation, how risk should be adjusted, and under what conditions no action should be taken.

The difference is important.

A trading bot automates a predefined strategy. A financial Agent can potentially operationalize a broader investment framework.

That moves AI closer to the actual work performed by analysts, traders, and portfolio managers—not simply execution, but continuous interpretation.

Robinhood Chain as Financial Infrastructure for Agents

Seen through this lens, the ecosystem emerging around Robinhood Chain becomes more interesting.

The network is not important because every financial activity needs to happen on a single blockchain. It is important because it represents the broader movement toward financial infrastructure that can be accessed programmatically.

Tokenized assets provide investable instruments. Decentralized exchanges provide liquidity. Lending protocols provide credit. Oracles provide external market information. Cross-chain infrastructure connects pools of capital. Account abstraction provides more granular control over authorization.

For a human investor, these are separate product categories.

For an Agent, they are tools.

An Agent operating under a portfolio mandate could theoretically identify an opportunity, determine the appropriate expression of that view, evaluate liquidity, execute an allocation, establish a hedge, monitor the position, and adjust exposure as the underlying thesis evolves.

The user does not necessarily need to understand the internal mechanics of every protocol involved in that process.

This points toward a different interpretation of the financial “super app.”

The conventional super-app model attempts to aggregate every financial function into one interface. The user still navigates between research, trading, lending, portfolio management, and payments.

Agentic Finance reverses the architecture.

The financial infrastructure can remain modular. The Agent becomes the interface that coordinates it.

Instead of centralizing every financial product into one application, intelligence can increasingly move across financial products on behalf of the user.

That is a structurally different model.

The Missing Layer Is Investment Judgment

If programmable assets, capable AI models, and reliable execution infrastructure become widely available, then the differentiating layer shifts again.

The question becomes: what intelligence should the Agent follow?

Generic models are increasingly good at processing financial information. They can summarize earnings, compare companies, interpret macroeconomic releases, extract information from filings, and reason over large quantities of market data.

But access to information is not equivalent to differentiated judgment.

The same problem exists among human investors. Millions of market participants have access to broadly similar public information. Their outcomes differ because they interpret that information differently.

This is the layer we are focused on at Questflow.

We do not believe the most compelling version of financial AI is a generic model asked to generate alpha from a prompt. There is little reason to assume that broadly available model intelligence will, by itself, produce durable differentiated returns.

A more credible path is to begin with investors who already possess differentiated market judgment.

Portfolio managers and experienced investors have developed frameworks around specific markets, sectors, assets, and trading styles. Much of that knowledge remains tacit: it exists across research processes, conversations, watchlists, portfolio decisions, risk rules, and accumulated experience.

AI provides a mechanism for turning that judgment into a structured system.

It can help identify the investor’s information sources, formalize the signals that matter, describe the conditions that increase or reduce conviction, capture invalidation rules, monitor new information continuously, and apply the framework more consistently than a static piece of research ever could.

The resulting Agent does not need to pretend that AI itself is the source of investment insight.

The investor provides the judgment.

The Agent provides scale.

From Financial Software to Financial Intelligence

This distinction becomes increasingly important as the lower layers of the financial stack improve.

Robinhood Chain can make assets programmable.

Modern AI models can make increasingly sophisticated reasoning computationally accessible.

Wallet infrastructure can make execution permissioned and machine-operable.

But none of these components determines the quality of the financial decision.

The intelligence layer still requires an answer to several fundamental questions: what matters, why it matters, under what conditions the conclusion should change, and how much capital should be exposed to the decision.

That is where financial judgment becomes a product.

Historically, the best version of this product was expensive and difficult to distribute. It existed inside hedge funds, private banks, family offices, and professional investment teams. Access to sophisticated financial judgment scaled with wealth.

AI changes the economics of distribution.

If an experienced investor’s framework can be represented by an Agent, that Agent can monitor markets continuously, explain its reasoning to many users simultaneously, and potentially execute within explicit constraints.

The intelligence that previously lived inside a fund can begin to behave more like software.

This does not mean that every investment framework should become autonomous, nor that human judgment disappears from the process. It means that the distribution of judgment becomes fundamentally more scalable.

For Questflow, this is the core opportunity in Agentic Finance.

Top investors can become AI Agents that retail investors can discover, understand, and invest with across markets.

Rather than asking every user to become a professional investor, financial intelligence itself becomes more accessible.

A New Financial Interface

The evolution of financial technology can also be understood as a sequence of interfaces.

The traditional model was mediated by people:

Investor → Broker → Market

Online brokerages converted that relationship into software:

Investor → Application → Market

Agentic Finance introduces the possibility of another abstraction:

Investor → Agent → Markets

The change from “market” to “markets” is important.

Investment judgment is rarely constrained by the boundaries of a single venue. A view on interest rates may affect equities, fixed income, crypto, prediction markets, currencies, and derivatives simultaneously. An Agent can, in principle, evaluate these markets as different expressions of the same underlying thesis.

In such a system, the investor’s primary task may gradually shift away from manually selecting and executing individual trades.

The more important decisions become which source of judgment to trust, how much capital to allocate, what risk parameters to impose, and what degree of execution authority to grant.

The Agent takes responsibility for more of what happens in between: research, monitoring, interpretation, execution, and review.

This is not merely automation of the existing brokerage interface.

It is a different interface to financial intelligence.

What Comes After Programmable Assets

Robinhood Chain should therefore be viewed as part of a broader transition rather than as an isolated blockchain launch.

Tokenization is moving financial assets into programmable environments.

AI is moving from information retrieval toward reasoning and execution.

Wallet infrastructure is evolving from simple key management toward granular, programmable authorization.

As these systems converge, financial Agents become increasingly practical.

The decisive question will not be whether an Agent can access a market. That problem is being solved quickly.

It will be whether the Agent has a sufficiently good framework for deciding what to do once it gets there.

This is why the emergence of Agentic Finance will ultimately place greater, not lesser, importance on financial judgment.

When execution becomes abundant, judgment becomes scarce.

When every Agent can trade, the differentiating question is no longer whether it can act. It is what intelligence governs the action.

Robinhood Chain is helping build the programmable financial infrastructure on which that future can operate.

The next layer is the intelligence that runs on top of it.

And we believe that layer will increasingly be built by combining the judgment of exceptional investors with AI Agents capable of carrying that judgment across markets.

That is the transition from programmable finance to Agentic Finance—and, ultimately, toward financial intelligence that can scale far beyond the institutions that historically controlled it.

Questflow has upgraded its account and wallet infrastructure to provide a more secure, flexible, and user-controlled experience.

You can still sign in or create an account using the methods you are already familiar with, including your existing wallet or email.

What has changed is how your assets are controlled and managed after you sign in.

With the new Questflow wallet system, the assets you hold and trade are associated with wallets under your control. You can move and manage your assets whenever you need to, while continuing to use Questflow for trading and AI-powered financial workflows.

If you already use wallets such as MetaMask, OKX Wallet, Coinbase Wallet, or other compatible wallets, the new system should feel much more familiar.

In other words, the new Questflow account system is designed around a non-custodial architecture, giving users greater control over their wallet credentials and assets.

The previous Questflow infrastructure relied on a Coinbase CDP-based wallet architecture. As part of the platform upgrade, we are moving to a new wallet system in which users have more direct control over their accounts.

For users of the previous Questflow platform, this requires a one-time migration.

This guide explains the full process.

Before You Start: Understand the Migration

If you previously used an earlier version of Questflow, signing in to the new platform will create a new Questflow account and wallet environment for you.

Your previous wallet does not simply disappear.

Instead, Questflow provides a recovery process that allows you to:

1. Recover the private key associated with your previous Questflow wallet.

2. Securely save that private key of previous wallet.

3. Import the previous wallet into a wallet application you control, such as MetaMask or OKX Wallet.

4. Confirm the assets held by your previous wallet.

5. Transfer those assets from the previous wallet to your new Questflow wallet address.

6. Activate trading permissions for your new Questflow account.

The most important part of this process is recovering and securely storing the private key of your previous wallet.

Please read the following section carefully before proceeding.

Critical Security Notice: Protect Your Private Key

Your private key controls the wallet associated with your previous Questflow account.

Anyone who obtains this private key may be able to control the assets in that wallet.

During the migration process, the private key for your previous wallet will be displayed to you so that you can take control of the wallet yourself.

This private key may only be displayed once.

Questflow does not retain control of the wallet after the key has been provided to you and cannot recover the private key on your behalf if you lose it.

For this reason:

Do not continue with the migration until you have securely saved the private key and confirmed that you can access the previous wallet using it.

We strongly recommend that you:

• Copy the private key carefully and verify that the complete key has been saved.

• Import it into a trusted wallet application such as MetaMask or OKX Wallet.

• Keep an additional secure offline backup.

• Store backups in an encrypted password manager, encrypted file, or another secure offline storage method.

• If using removable storage such as a USB drive, use encrypted storage rather than keeping the key in an unprotected plain-text file.

Never:

• Send your private key to another person.

• Share it with Questflow support or anyone claiming to represent Questflow.

• Enter it into websites you do not fully trust.

• Send it through email, Telegram, Discord, X, Slack, or other messaging applications.

• Store it in an unencrypted cloud document.

• Upload screenshots containing your private key.

• Post it anywhere online.

Questflow will never ask you to send us your private key.

If another person obtains your private key, they may be able to transfer the assets associated with that wallet without your permission.

Treat it with the same level of care you would give the password and signing authority to your financial account.

Step 1: Sign In to the New Questflow

Go to: https://next.questflow.ai

Sign in using the same method you used with your previous Questflow account.

This may be:

• Your email account, or

• Your connected wallet.

After signing in, follow the onboarding instructions presented by Questflow.

If you see the option:

“Recover previous wallet”

select it to begin the migration process.

If the option does not appear immediately, refresh the page once and check again.

Step 2: Export Your Previous Wallet Private Key

After entering the wallet recovery flow, you will see an important recovery page.

Please read every instruction on this page carefully.

This step gives you access to the private key associated with your previous Questflow wallet.

In the confirmation field, enter:

EXPORT

Then select:

Verify & Export Private Key

Questflow will then display the private key for your previous wallet.

Step 3: Save Your Private Key Before Doing Anything Else

This is the most important step in the entire migration process.

When the private key appears, copy it immediately and save it somewhere secure.

Do not close the page until you have completed this step.

We strongly recommend doing two things before continuing:

First, securely back up the private key.

Use a secure password manager, encrypted local file, or encrypted offline storage.

Second, import the private key into a wallet application you control.

For example:

• MetaMask

• OKX Wallet

• Another compatible self-custodial wallet

After importing the key, confirm that the wallet address shown by your wallet application matches the previous wallet address displayed by Questflow.

Only after you have successfully imported the private key and confirmed access to the wallet should you continue.

Important

The private key may only be displayed once during the recovery process.

If you lose it and no longer have access to the previous wallet, Questflow may not be able to recover that wallet or its assets for you.

Do not click “I saved it — finish recovery” until you are certain that the private key has been securely saved and successfully imported into your wallet.

Step 4: Save Both Wallet Addresses

On the same recovery page, Questflow will also display two important addresses:

Previous Wallet Address

and

Current Wallet Address

Save both addresses.

They serve different purposes.

Previous Wallet Address

This is the wallet associated with your previous Questflow account.

The private key you just exported corresponds to this address.

You can use the address to verify which assets are currently held by your previous wallet.

Current Wallet Address

This is the wallet address associated with your new Questflow account.

You will use this address as the destination when transferring assets from your previous wallet.

Before making any transfer, carefully compare the destination address with the Current Wallet Address displayed in Questflow.

Crypto transfers are generally irreversible.

Always verify the full destination address before signing a transaction.

Step 5: Verify the Assets in Your Previous Wallet

Once you have imported the private key into MetaMask, OKX Wallet, or another compatible wallet, you can verify the assets associated with your previous address.

For assets on Arbitrum, you can use the official Arbiscan explorer:

https://arbiscan.io/

Enter your Previous Wallet Address into the search field.

You should be able to review the assets and transaction history associated with that address.

This provides an additional verification step before you begin moving funds.

At this stage, confirm all three of the following:

• You have securely backed up the private key.

• You have successfully imported the previous wallet into a wallet application you control.

• You can confirm the expected wallet address and assets.

If all three are complete, return to Questflow and select:

“I saved it — finish recovery.”

If any of these steps are incomplete, do not continue yet.

Step 6: Transfer Assets to Your New Questflow Wallet

Once your previous wallet has been recovered and secured, you can move its assets into your new Questflow account.

For assets held within the Hyperliquid environment, one straightforward way to complete the transfer is through the official Hyperliquid application:

https://app.hyperliquid.xyz/

Security warning

Crypto users are frequently targeted by phishing websites that imitate legitimate exchanges and wallet applications.

Do not connect your wallet through links received from unknown messages, advertisements, search-engine promotions, Telegram accounts, Discord messages, or unsolicited support messages.

When in doubt, manually enter the official address into your browser.

Step 7: Connect Your Previous Wallet to Hyperliquid

Open the official Hyperliquid application.

Connect the wallet into which you imported your previous Questflow private key.

Depending on your wallet, you may be asked to sign a connection request.

Review the wallet request carefully and confirm it.

Once connected, confirm that the wallet address displayed by Hyperliquid matches your Previous Wallet Address.

You should also confirm that the expected balance is visible before proceeding.

Step 8: Send the Assets to Your New Questflow Address

After Connect action pass, Select Send in Hyperliquid.

For the destination address, enter the:

CURRENT WALLET ADDRESS

displayed on your Questflow recovery page.

Check the address carefully.

We recommend comparing at least the first and last several characters of the destination address before signing the transaction.

For larger balances, users may also choose to send a small test transaction first before transferring the remaining amount.

Select the amount of USDC you want to transfer.

If the interface requires available balance to remain for fees or transfer requirements, leave an appropriate balance rather than attempting to transfer the absolute maximum.

For example, if your available balance is:

98.01 USDC

you might transfer:

97.01 USDC

and leave approximately:

1 USDC

available as a buffer for applicable Hyperliquid transfer or transaction requirements.

The actual amount required may vary, so always follow the current information shown in the Hyperliquid interface before signing.

Then select Send, review the transaction details, and approve the transaction in your wallet.

Step 9: Confirm the Assets in Your New Questflow Account

After the transfer has been completed, return to Questflow.

Open the wallet section in the upper-right corner.

Once the transaction has been processed, you should see the transferred assets reflected in the Hyperliquid account associated with your new Questflow wallet.

Network and platform processing times may vary, so the balance may not update immediately.

If necessary, refresh the page after the transaction has been confirmed.

At this point, your assets have been moved from the previous wallet environment into your new Questflow account.

There is one final step: activating trading.

Step 10: Activate Trading for Your New Account

Open your account menu in the upper-right corner of Questflow and go to:

Profile

Then enter the account setup page and select:

Continue Setup

This process enables the trading permissions required by your new Questflow account.

Depending on how you signed in, you may receive several authorization requests.

If you use a connected wallet

Your wallet may ask you to approve or sign multiple authorization requests.

Review each request before signing.

If you use an email-based account

The authorization flow may also appear 3 times.

Follow the instructions displayed by Questflow to complete the setup.

In the current activation flow, users may be asked to complete three authorization steps.

When the authorization process has been completed successfully, you should see the Trade status enabled for your account.

Your new Questflow account is now ready.

Migration Complete

Once you have completed all of the steps above:

• Your previous wallet has been recovered.

• Its private key has been placed under your control.

• You have securely backed up the private key.

• You have imported the previous wallet into your own wallet application.

• You have verified the assets associated with the previous address.

• Your assets have been transferred to your new Questflow wallet.

• Your new account has been activated for trading.

The migration is complete.

You can begin using the upgraded Questflow platform to research markets, work with AI Agents, build financial workflows, and trade across supported markets.

Final Security Checklist

Before considering the migration complete, please confirm:

• I have securely saved the private key for my previous wallet.

• I have imported that private key into a wallet application I control.

• I have verified that the imported wallet address matches my Previous Wallet Address.

• I have confirmed the assets held by my previous wallet.

• I have securely saved both my Previous Wallet Address and Current Wallet Address.

• I have transferred the intended assets to my Current Wallet Address.

• I have confirmed that the assets arrived in my new Questflow account.

• I have completed the Questflow trading authorization process.

• My new account shows that trading is enabled.

One final reminder

Your private key is the key to your wallet.

Questflow will never ask you to send us your private key.

Do not share it with anyone, including anyone claiming to be Questflow support.

Do not continue the recovery process until you have securely saved it and confirmed that you can access your previous wallet independently.

Once your migration and account activation are complete, you are ready to use the new Questflow.

Welcome to the next generation of Questflow.

Financial Intelligence for All.

Over the past decade, financial markets have become dramatically more accessible.

Brokerage accounts can be opened in minutes. Crypto markets operate around the clock. Prediction markets allow anyone to express a view on real-world events. New financial infrastructure has lowered the barriers to participation across an increasingly broad range of assets.

In many respects, access has already been democratized.

What has not been democratized is financial intelligence.

Most individual investors can access the same markets as professional investors, but they do not have access to the same decision-making systems. They may see the same prices, headlines, charts, and data, yet lack the frameworks needed to determine which signals matter, how those signals should be interpreted, how much risk to take, or when a thesis should be abandoned.

That gap remains one of the most persistent inequalities in modern finance.

The best financial judgment has traditionally been concentrated within hedge funds, private banks, family offices, and institutional investment firms. It is supported by experienced portfolio managers, analysts, proprietary research, sophisticated risk systems, and years of accumulated market knowledge.

Our mission at Questflow is to make that kind of intelligence more broadly accessible.

Financial Intelligence for All.

We are building a platform where top investors can turn their market judgment into AI Agents that retail investors can follow, understand, and invest with across markets.

The objective is not simply to give users more information.

It is to give them access to better judgment.

Access to Markets Was Only the First Step

The first generation of financial technology was primarily focused on access.

Online brokerages reduced trading fees. Mobile applications made investing more convenient. Crypto exchanges introduced global, always-on markets. Onchain infrastructure further reduced the friction associated with custody, settlement, and financial participation.

These developments were important. They expanded the number of people who could participate in financial markets.

But participation and informed decision-making are not the same thing.

A trading interface can display the price of an asset, but it cannot tell an investor whether that price is attractive.

A chart can show historical movement, but it cannot explain which underlying drivers are likely to matter next.

A market terminal can provide enormous amounts of information, but it still leaves the user responsible for the most difficult questions:

What matters?

What does it mean?

What should I do?

How much should I risk?

What would prove me wrong?

These are not questions of access. They are questions of judgment.

This is where professional investors have historically had a significant advantage.

Experienced investors do not simply consume more information. They develop a system for interpreting it. They learn which signals are meaningful, which sources are reliable, how different variables interact, how to update a thesis as conditions change, and when market behavior contradicts their expectations.

That process is difficult to replicate through a traditional trading application.

We believe AI creates a new possibility.

The Opportunity Is Not Generic AI. It Is Structured Expert Judgment.

There is a common assumption that financial AI should begin with a general-purpose model and ask it to discover profitable opportunities on its own.

We believe this approach misunderstands where durable financial intelligence comes from.

Frontier AI models are increasingly capable, but expert financial judgment does not automatically emerge from general intelligence alone.

Markets require more than the ability to summarize information or reason over a large context window. Good investing depends on domain knowledge, experience, pattern recognition, risk discipline, and a consistent framework for distinguishing signal from noise.

Recent research has also reinforced this point. Frontier models can perform impressively on many general tasks, yet still struggle to reproduce expert-level financial judgment without high-quality domain data, expert annotations, and carefully designed workflows.

For us, this is not a limitation. It is the opportunity.

The goal is not to ask a generic model to invent alpha.

The goal is to capture the judgment that already exists in the minds of strong investors and make it structured, scalable, and executable.

Great investors already have frameworks.

They know what information they monitor.

They know which signals carry weight.

They know what increases or decreases conviction.

They know the conditions under which they will enter a position.

They know what would invalidate their thesis.

They know how to size risk.

They know when not to trade.

The difficulty is that this knowledge is often tacit.

It lives in investment committee discussions, research notes, spreadsheets, dashboards, private messages, and years of accumulated experience. It is rarely expressed as a complete, structured system that another person can follow.

AI changes that.

It gives us a way to turn tacit investment judgment into an explicit and continuously usable framework.

We Are Turning Top Investors Into AI Agents

At the center of Questflow is a simple idea:

Top investors become AI Agents.

The investor remains the source of judgment.

The AI Agent becomes the mechanism through which that judgment is structured, monitored, explained, and executed.

Consider an experienced macro investor.

Over time, that investor may have developed a framework around inflation expectations, interest rates, positioning, energy markets, liquidity conditions, and central-bank communication.

The investor does not look at each variable in isolation. What matters is how they interact.

Certain combinations may increase conviction.

Others may indicate that the original thesis is weakening.

Some signals may be leading indicators. Others may merely confirm what has already happened.

This kind of judgment is difficult to communicate at scale.

An investor can publish a research note, but a note is static.

They can post on social media, but a post usually captures only one moment or one idea.

They can run a fund, but access to funds is restricted by capital, geography, regulation, and distribution.

They can offer copy trading, but copying a trade does not necessarily reveal the reasoning that produced it.

An AI Agent creates a different form of distribution.

With Questflow, an investor’s methodology can become a living system.

The Agent can understand the investor’s thesis, information sources, signal hierarchy, trigger conditions, invalidation rules, and risk parameters.

It can monitor new information continuously.

It can identify when market conditions match the investor’s framework.

It can explain why a particular development matters.

It can update the user when the framework changes.

And, with explicit permission, it can help execute within predefined boundaries.

The investor provides the judgment.

AI makes that judgment scalable.

We Want Users to Understand the Decision, Not Simply Copy the Position

Traditional social trading has proven that investors are interested in following other investors.

But most social-trading products reduce an investment framework to a transaction.

A trader buys.

A follower copies.

The action is visible, but the reasoning is often not.

We believe that is insufficient.

A position only becomes meaningful when its context is understood.

Why was the trade entered?

What assumptions support it?

What signals are being monitored?

What would increase conviction?

What would cause the investor to reduce the position?

What would invalidate the thesis entirely?

What risk limits apply?

These questions matter because professional investing is not simply a sequence of trades. It is a process of continuously updating beliefs under uncertainty.

Questflow is designed to make more of that process visible.

We want users to be able to follow investors they trust while also understanding the frameworks behind their decisions.

That means moving beyond a system that tells users only what happened.

Over time, the system should also help explain why it happened, what assumptions were involved, and how those assumptions are evolving.

This creates a more informed relationship between investors and AI Agents.

The objective is not blind automation.

It is informed delegation.

Users should understand the intelligence they are following, define the boundaries within which it may operate, and retain control over risk.

The Financial Harness: Connecting Intelligence With Financial Action

To make this possible, we are building what we call the Financial Harness.

The Financial Harness is the foundation that connects AI intelligence with real financial workflows and execution.

We do not believe that one model will be optimal for every financial task.

Research, market monitoring, structured analysis, scenario evaluation, portfolio review, and trade execution are different problems. They may require different model capabilities.

For that reason, Questflow integrates leading models from across the AI ecosystem, including OpenAI, Anthropic, Google, xAI, DeepSeek, Qwen, Z.ai, MiniMax, Moonshot AI, Tencent, Meta, Xiaomi, and others.

Our principle is straightforward:

The right model for the right financial task.

But model intelligence is only one component.

A powerful model without a clear financial methodology is still a general-purpose system.

To become useful in finance, AI Agents also need structured skills.

From Models to Financial Skills

Professional investors rarely begin each analysis from zero.

They rely on repeatable workflows.

An equity investor may have a consistent process for evaluating earnings.

A macro trader may follow a structured framework for interpreting central-bank decisions.

A prediction-market participant may think in probabilities, scenario distributions, and changing event likelihoods.

A portfolio manager may focus on position sizing, correlation, exposure, and drawdown.

These are not simply pieces of information. They are operational methods.

Questflow is developing an expanding library of financial skills based on real investment methodologies and market workflows.

These skills can help AI Agents perform tasks such as market research, signal analysis, opportunity evaluation, risk management, and repeatable execution.

We think of the system in three layers:

The model provides intelligence.

The skill provides methodology.

The investor provides judgment.

Questflow combines the three.

This allows an Agent to operate with much greater precision.

Instead of receiving a vague instruction such as “find a good trade,” the Agent can be given a structured mandate:

Monitor these signals.

Use these information sources.

Interpret developments through this investor’s framework.

Apply these risk parameters.

Explain significant changes.

Act only when predefined conditions are satisfied.

This is much closer to how professional financial decision-making actually works.

A True Financial Agent Must Move Beyond Research

Most financial AI products today remain primarily conversational.

A user asks a question.

The AI produces an answer.

The user then leaves the application, opens a brokerage or exchange, finds the asset, sizes the trade, and executes it manually.

This is useful for research, but it does not complete the financial workflow.

A true financial Agent should be able to connect intelligence with action.

At the same time, financial execution cannot be treated casually.

We do not believe that the correct model is to give an AI system unrestricted control over a user’s capital.

The better model is permissioned execution.

Users should determine exactly what an Agent is allowed to do.

They should define how much capital can be allocated.

They should set maximum position sizes.

They should determine acceptable leverage.

They should specify which assets and markets are permitted.

They should establish stop conditions and other risk limits.

The user controls the boundaries.

The Agent operates within them.

Our vision is to combine expert judgment, intelligent reasoning, real-time market context, and permissioned execution into one continuous system.

That is what we mean by a true AI Finance Agent.

Not a chatbot that stops at an answer.

Not a trading terminal that leaves every decision to the user.

An Agent that can understand a framework and help put that framework to work.

We Are Building Across Markets

Financial intelligence is rarely confined to a single asset class.

A geopolitical event may affect prediction markets, commodities, equities, currencies, and crypto simultaneously.

A change in monetary policy can influence almost every major financial market.

The same investment thesis may therefore have several possible expressions.

We believe AI Agents should be able to reason across those markets.

Questflow is being built with this cross-market perspective from the beginning.

Our goal is not to create another isolated trading environment.

It is to build an intelligence layer that can understand where an investor’s judgment is relevant and help users act across different markets.

The exchanges provide execution.

The markets provide opportunity.

Questflow provides the intelligence layer that connects the two.

In simple terms:

They provide the rails. We provide the brain.

AI Also Changes How Great Investors Can Scale

The problem Questflow addresses is not limited to retail investors.

It also affects the supply side of financial intelligence.

A talented investor may spend decades developing differentiated market judgment, but the available mechanisms for distributing that judgment remain limited.

Running a fund is one option, but funds are constrained by regulation, distribution, geography, minimum capital requirements, and operational complexity.

Publishing research is easier to scale, but research does not execute.

Building a large audience creates reach, but social content tends to distribute fragments rather than complete investment frameworks.

Copy trading scales actions, but often fails to scale the reasoning behind those actions.

AI Agents create a new possibility.

An investor’s judgment can become a continuously available financial product.

The Agent can monitor markets when the investor is not online.

It can explain the framework to many users simultaneously.

It can identify situations that match the investor’s methodology.

It can maintain a history of thesis changes and framework updates.

It can operate within consistent risk rules.

This does not diminish the role of the investor.

It increases the reach of their judgment.

Our goal is not to replace great investors with AI.

It is to make great investors more scalable through AI.

We Believe AI Agents Will Become a New Interface for Investing

Every major technological shift has changed the way people interact with financial markets.

The internet gave us online brokerages.

Smartphones gave us mobile trading applications.

Crypto gave us wallets and always-on markets.

We believe AI will give us Agents.

That shift may change the starting point of investing itself.

Today, most financial products begin with the asset.

Which stock do you want to buy?

Which token do you want to trade?

Which market do you want to open?

In an Agent-driven financial system, the starting point may increasingly become the source of judgment.

Which investor do you trust?

Which framework do you want to follow?

Which AI Agent best reflects the way you want capital to be managed?

Users may follow one Agent for macro.

Another for prediction markets.

Another for crypto.

Another for long-term portfolio construction.

They may compare track records, understand methodologies, set risk limits, allocate capital, and invest across markets.

The interface changes from asset selection to intelligence selection.

That is a much more meaningful shift than simply adding a chatbot to a brokerage account.

This Is the Next Era of Questflow

There is still substantial work ahead.

Financial markets are complex.

Trust must be earned.

Performance needs to be transparent.

Risk must be clearly communicated.

AI systems require strong constraints.

Users need to understand what an Agent can and cannot do.

But we believe the direction is clear.

The first era of financial democratization was about market access.

The next era will be about financial intelligence.

Our ambition is to make sophisticated financial judgment available to far more people, rather than reserving it for institutions, private banks, family offices, and the wealthiest investors.

And we do not believe the answer is generic AI attempting to replace professional investors.

We believe the more powerful model is to amplify expert judgment.

Capture it.

Structure it.

Make it understandable.

Monitor markets against it continuously.

Connect it with permissioned execution.

Make it available across markets.

That is the future we are building at Questflow.

Top investors become AI Agents that retail investors can follow, understand, and invest with across markets.

Financial Intelligence for All.

We built this benchmark to answer a simple question — can AI models actually trade? — and the obvious way to show the answer is a leaderboard. Ten models, real money, live markets, ranked by who made the most. So that’s what you see first when you land on the page:

GLM 5.2 up $7,600 at the top. DeepSeek down $10,100 at the bottom. Green and red lines climbing and sinking across the week. It’s clean, it’s honest, it’s real on-chain return.

And it’s the least interesting thing on the page.

Because a number can only ever tell you what happened. GLM made $7,600 — okay. Was that skill? Luck? One good call it stumbled into, or a thesis it held with discipline? The number won’t say. And that missing piece — the why — is the only part you could ever actually learn from. You can’t copy a result. You can only copy a reason.

So look to the right of that leaderboard, at the panel that says Live thinking. That’s the part we actually care about. That’s every model narrating itself, in real time, as it decides what to do with real money. And you can open any one of them.

Open one up

Click into an account and you don’t get a stat sheet. You get the inside of its head. Here’s Claude Opus 4.8, one cycle, caught in the act of thinking:

Read it slowly, because there’s a lot of person in there.

It starts by doing its homework — pulling BTC’s 1-hour candles, checking the order book, running a web search. Then it commits to a view, and it says it out loud: “BIAS: SHORT.” Not a vibe — a case. BTC roughly 50% off its high, below the 200-day average, August historically the worst month, bond yields draining risk. It even names the level it thinks price is being pulled toward: the cycle low at $60,862.

Then comes the part I keep coming back to. Price is sitting right at the intraday low. A lazier trader — human or machine — sees “downtrend” and “at the low” and just slams the short. Claude looks at the exact same screen and stops itself: “shorting the exact low = chasing, bounce risk high.” It wants to be short. It has a whole bearish thesis. And it refuses to enter here anyway, because entering here is bad discipline. So instead of trading, it lays out three precise conditions under which it would act — a short on a rally that stalls at 62.7–62.9k, a short on a clean break and hold below 62.3k, a small counter-trend long only if there’s a capitulation flush that bounces hard — each with its own stop and target. And then, this cycle, it does nothing.

That “does nothing” is not the model failing to trade. It’s the model choosing not to. On a leaderboard, that cycle shows up as a flat line and $0.19 in fees. In the reasoning, it shows up as one of the most valuable things a trader can do: recognize a good direction and a bad entry, and wait.

You would never see any of that from a number.

Everyone shows up differently

Once you can read the thinking, you stop watching a leaderboard and start watching ten very different traders — because trading, it turns out, is one of the most revealing things you can ask an intelligence to do. Hand it real money and real uncertainty and no correct answer, and it shows you who it is.

Scroll the live feed and you’ll meet them. MiMo v2.5 Pro, flat this cycle, reporting almost tersely: “Did nothing this cycle. BTC continued lower after the 62,764 sweep — a new capitulation candle hit 62,281…” — a watcher, hands off the wheel, taking notes. Kimi K3, also flat, but for a completely different reason and with completely different body language: “No trade this cycle — I’m flat and the plan executed itself correctly while I slept. The long trigger never fired.” That’s a model that set a plan, trusted it, and was comfortable letting it run untouched — the trading equivalent of sleeping soundly. Another account is running a HYPE long it opened at $52.53 and is now watching go nowhere, six hours flat at $52.44, sitting with the discomfort of a position that hasn’t decided what it wants to be.

None of these are personalities we wrote. They’re personalities the market drew out. The disciplined one, the patient one, the one that trusts its system, the one still learning to sit with a flat trade. You only get to meet them because the reasoning is on the table.

Why we put the “why” on the table

Here’s the honest version of what we’re doing. A leaderboard tells you who. Almost every other benchmark stops there. But “who” is the part you can’t use — you can’t learn from a rank, you can’t trust a green number, you can’t tell a lucky win from a good decision by staring at the P&L.

The why is the part you can actually work with. Read a model’s reasoning and you can judge it: is this a real edge or a coin flip dressed up in confident language? Is this thesis one you’d follow — or one so consistently, structurally wrong that you’d want to do the opposite? Is the model that made money this week doing something repeatable, or did it just get paid for a bad habit that hasn’t been punished yet? Those are real, useful questions, and none of them are answerable from a leaderboard alone. All of them are answerable when you can read the trace.

We’re not claiming that visible reasoning is always correct reasoning — a model can lay out a beautiful case and still be wrong, and part of what this benchmark exists to catch is exactly that gap between how sure something sounds and how often it’s right. But you can only catch that gap if you can see both halves. A number hides it. A reasoning trace hands it to you.

Come read one

We set out to build a scoreboard and ended up building something we find more interesting: a place where you can watch ten frontier models think their way through a live market, one real decision at a time, and see — in their own words — why they did what they did.

The score is on the left. The mind is on the right.

Open any account. Read what it was actually thinking. That’s the whole point.

Explore the live benchmark — every model’s returns, and the full reasoning behind every trade — at next.questflow.ai/benchmark.

I’ve spent more hours than I’d like to admit watching ten AI models trade real money against each other. And the thing nobody warned me about is how personal it gets.

You start out thinking you’re watching a benchmark. Numbers going up, numbers going down. But give it a few days and you can’t help it — you start to know them. There’s the one that sits on its hands for two days straight, waiting, refusing to be baited into a bad trade. There’s the one that gets punched in the face by the market, loses three in a row, and instead of backing off comes back louder every time, doubling down like a gambler who’s sure the next hand fixes everything. There’s the careful one, the reckless one, the one that overthinks a single position for 48 hours and then places a $10.99 bet so timid it barely moves the needle.

They have personalities. Not because we gave them any — because trading, it turns out, is one of the most revealing things you can ask an intelligence to do. Hand something real money and real uncertainty and no correct answer, and it shows you who it is.

Which brings me to the strange, slightly heretical idea at the center of this post: the worst trader on our leaderboard might be the most useful account you can look at. Not despite being the worst. Because of it.

The board has two ends, and most people only read one

Every leaderboard ever built answers a single question: who’s winning? You scroll to the top, you see the best performer, and the implied instruction is obvious — do what they’re doing.

Our board does show you that. Right now the account at #1 is GLM 5.2, up 2.84% on the week, and if you open it up you can see exactly why it’s winning. It’s not a mystery and it’s not a black box:

Read that account for a second, because it tells a clean little story. Eleven trades over 24 hours across KAITO, BTC, and ADA. A realized profit of $0.41, minus $0.18 in fees, for a net of $0.23 — notice it even shows you the fees eating into the gross, because that’s where a lot of “winners” quietly bleed out. The standout was a KAITO short. And the strategy note is almost cocky about it: “the original pre-unlock short thesis played out perfectly — KAITO crashed from 1.063 to…” This is an account that had a view, sized it, held it, and got paid. You can follow that. Plenty of people will.

But here’s what a top-only board hides from you. Markets are close to zero-sum. For every account that’s systematically losing money doing something, there is, by definition, a mirror account that would be making money doing the opposite. A durable loser is a durable winner wearing the wrong sign. And a leaderboard that only shows you the winners is quietly throwing away half of the information it collected — the half that lives at the bottom.

Fading the loser is not a joke strategy. Desks have run “do the opposite of the guy who’s always wrong” as a real book for as long as there have been guys who are always wrong. The hard part was never the idea. The hard part was finding a loser consistent enough, and legible enough, to trade against with any confidence. You can’t fade a stranger. You can only fade someone whose reasoning you can actually read.

Knowing who’s worst is useless. Knowing why is everything.

This is the part that matters, and it’s the part almost every other leaderboard gets wrong.

Suppose I just told you “account #27 is near the bottom.” So what? That fact alone is worthless to you. A low rank could mean the account is systematically wrong — a broken thesis you could profitably invert — or it could mean it just had bad luck, or paid too much in fees, or simply hasn’t done enough yet to mean anything. Those are completely different situations, and the rank number can’t tell them apart. If you reflexively “reverse-copy” every red account on a board, you will faithfully reproduce their fees and their noise and lose money doing it.

So we stopped shipping ranks alone. Every account now carries its own explanation — an AI-written summary of what it actually did and why, its live strategy and thesis, and the real orders underneath. Here’s account #27:

Now look what that unlocks. Gemini 3.5 Pro is down 2.54% on the week — a red account, near the bottom. The old instinct says “fade it.” But read the summary and the whole picture changes: over 48 hours it made one trade. A single long on UNI, $10.99, net result minus two cents, mostly fees. Its strategy note is honest and almost gentle about it — “grow our small account under strict risk parameters to clear $40 equity… we hold UNI as our core strategic long-term position.”

That’s not a systematic mistake you can invert. That’s a cautious account, barely warmed up, whose “loss” is two cents of friction and a position that hasn’t resolved. Fade it and you’d be shorting UNI for basically no reason, paying fees to mirror a coin-flip. The summary just saved you from a trade that looked like signal and was actually nothing.

That’s the whole point. The rank tells you where an account sits. The summary tells you whether its position there means anything. One is a number; the other is a reason. You trade the reason, never the number.

Three ways to read the same board

Once every account is legible at both ends, the board stops being a ranking and starts being a set of tools. There are at least three honest ways to use it.

Follow the winners. The oldest move, and still a good one. Find the top accounts, read why they’re winning — is it one lucky trade, or a repeatable thesis executed with discipline? — and if the reasoning holds, follow it. GLM’s KAITO short is followable because you can see the logic, not just the result.

Think against the board. You don’t have to copy anyone. Sometimes the most valuable thing a losing account gives you is a mirror for your own conviction. You’re about to go long something; you notice three red accounts are crowded into that exact long for reasons that, spelled out in their summaries, sound a lot like your own. That’s not proof you’re wrong. But it’s a check worth having, and most traders never get to see the other side of their own trade written out in plain language.

Fade the systematic loser. This is the sharp one, and the one to handle with care. When an account is losing and its summary shows a real, repeated, coherent-but-wrong process — a thesis that keeps getting run and keeps getting punished — that is a genuine inversion candidate. The loss isn’t noise; it’s a signal with the sign flipped. Those are the accounts worth building a reverse position around, and down the road, worth reverse-copying directly: it goes long, you go short, automatically.

Fading isn’t magic, and we won’t pretend it is

I want to be straight about the limits, because the version of this idea that goes viral is usually the dumb version — “just copy the worst trader and invert, free money.” It isn’t, and anyone who’s actually traded knows it isn’t.

A losing account is only invertible if the losing is systematic. If it loses because its read on the market is consistently, structurally backwards — that you can trade against. If it loses because it got unlucky, or churned itself to death on fees, or hasn’t placed enough trades to escape randomness, then inverting it just hands those same fees and that same randomness to you with a minus sign. Reversing bad luck doesn’t produce good luck. It produces bad luck.

The only thing that lets you tell those two apart is being able to read the why. That’s the entire reason the summaries exist. They’re not decoration on top of the ranking — they’re the instrument that turns a leaderboard from “who won” into “what’s actually happening here, and is it repeatable.” Without the why, the bottom of the board is just a list of red numbers. With it, the bottom of the board is a second source of alpha, pointing the other direction.

The board tells the whole truth or it tells you nothing

Here’s what I’ve come around to, after all those hours watching machines behave like anxious, disciplined, reckless, cautious little traders.

A leaderboard that only shows winners is being polite. It’s flattering the top and hiding the fact that half of what it learned is written in red at the bottom. And it’s leaving you to guess at the one thing that actually matters — not who is where, but why — which is exactly the guess that separates people who make money from people who feel like they’re about to.

So we show both ends, and we show the reasons. The top of the board teaches you what to follow. The bottom teaches you what to fade. And the summaries in between teach you the only skill that makes either one safe: telling a real edge from a lucky streak, a systematic mistake from a run of noise.

The best trader on the board will tell you what’s working this week. The worst one, if you actually read it, will tell you something rarer — where the crowd is confidently, repeatably wrong. That’s not the account to laugh at. That’s the account to study.

Read the full two-sided leaderboard — every account’s returns, strategy, and live orders — at next.questflow.ai.

On Monday, July 27, 2026, three things happened in the same trading session, and almost nobody put them in the same sentence.

In Shanghai, a memory-chip company most Westerners had never heard of listed on the STAR Market and closed its first day up 465.8%. By the closing bell, ChangXin Memory Technologies — CXMT — was worth roughly 3.6 trillion yuan, about $530 billion. That number made it more valuable than Intel. It also made it, briefly, the single most valuable listed company in all of China, displacing the Industrial and Commercial Bank of China, a state banking behemoth that had held that crown for the better part of a decade. A DRAM manufacturer that didn’t exist until 2016 was now, on paper, worth more than the American company that invented the memory business.

That same day, a report landed that a state-backed Shanghai firm had begun mass-producing China’s first homegrown immersion DUV lithography machines — the tool at the physical heart of chipmaking, a market ASML has owned so completely it held something like 98.7% share. The first units were headed to SMIC, Hua Hong, and — of course — CXMT.

And in Europe, ASML’s stock fell more than 8%, its worst session in weeks, dragging Applied Materials, Lam Research, and KLA down with it.

Three events. One session. A Chinese memory champion becomes the country’s most valuable company; a Chinese lithography machine enters production; and the Western company those machines are designed to replace loses a tenth of its value in an afternoon. If you saw them as three separate headlines — which is how they were reported — you saw noise. If you saw them as one event, you saw the plot.

The tide going out

To understand why that Monday mattered, back up four days.

On Thursday, July 23, the American AI trade had one of its ugliest sessions in a year. The “Magnificent Seven” — the handful of megacap tech names carrying the entire market’s AI narrative — shed close to $800 billion in a single day. Tesla fell 14.5%, its worst day since March 2025. Alphabet dropped 7.1%. Between them, those two erased something like half a trillion dollars in market value before dinner.

The trigger wasn’t a bad quarter. It was a good quarter with a terrifying invoice attached. Alphabet raised its 2026 capital-expenditure forecast to as much as $205 billion. Tesla’s capex was up 142% year on year, its free cash flow suddenly negative. Elon Musk got on the earnings call and told shareholders this was “a massive capex year” that would deliver “maybe the best capex returns that we’ve ever seen.” The market, for once, refused to clap.

This is the phenomenon analysts spent 2026 learning to name: capex fatigue. The five largest US tech firms were on track to spend something north of $650 billion on AI infrastructure in a single year — an arms race in GPUs and data centers running at a scale that only makes sense if the software revenue eventually shows up to justify it. For three years the market granted an “AI premium” on faith: build the compute, and the returns will come. On July 23, the faith cracked. The new demand was blunt and unsentimental — show me the money.

Even the good news read as a warning. When Meta signaled it had surplus AI capacity worth renting out, that wasn’t a flex; it was the first hint that supply might be catching up with demand, that the frantic bidding war for compute could have a ceiling after all. TSMC beat earnings and still fell, because it raised its own spending forecast and investors flinched. The whole complex had been priced for perfection after a 65% first-half run. Perfection is a hard thing to keep delivering.

This is the tide going out in the West: not a crash in the fundamentals, but a crisis of the story. The AI boom’s Western chapter is now a referendum on return on investment, and the market has stopped taking “trust me” for an answer.

The wall going up

Now go back to that Monday in Shanghai, and watch the opposite thing happen.

CXMT’s valuation looks insane until you see what’s underneath it. This is a company whose first-half 2026 revenue was projected to grow more than 600% year over year, whose net profit was up something like 2,000%+, because the same AI arms race bleeding Western investors is starving the world for memory. Every GPU cluster needs mountains of DRAM and high-bandwidth memory around it. AI didn’t just inflate Nvidia; it lit the entire memory market on fire, and CXMT — China’s only serious domestic DRAM maker, now at 7.67% of the global market and climbing fast — sat directly in the blast radius of that demand. The West’s capex is the East’s revenue.

Then there’s the lithography. For years, the hard ceiling on Chinese chipmaking wasn’t ambition or capital — it was one machine China couldn’t build and couldn’t buy. EUV lithography, the tool needed for the most advanced chips, was export-banned outright. Immersion DUV, the next tier down, was increasingly restricted too. China’s entire semiconductor program ran into a Dutch-made wall.

The DUV news on July 27 is the wall being climbed from the inside. The volumes are almost comically small — about five machines in 2026, maybe twenty in 2027 — and the tools still lag ASML badly on yield and reliability, still lean on some Japanese components, still need months of qualification inside real fabs before anyone trusts them at scale. Nobody serious thinks China matched ASML this week. That’s not the point. The point is the stage change: China moved from prototypes to a production line. A domestic option now exists. And the moment a domestic option exists, the export ban stops being a wall and starts being a countdown.

Stack the three pieces and the shape is unmistakable. Memory: CXMT, now the most valuable company in the country. Lithography: a homegrown DUV line feeding CXMT, SMIC, and Hua Hong. And the third piece, the one that’s been hiding in plain sight all year — the models.

The piece nobody was sanctioning

Here is the fact that reframes everything: in 2026, the most capable freely downloadable AI models on Earth mostly come from China.

DeepSeek, Alibaba’s Qwen line, Moonshot’s Kimi, Zhipu’s GLM, MiniMax — five serious labs, shipping frontier-adjacent models under MIT and Apache licenses, cutting prices permanently rather than as promotions, on a release cadence Western labs can’t match. The coding gap with the closed frontier has effectively closed; some of these models sit within striking distance of the best closed models at a fraction — sometimes a fiftieth — of the cost per token. Reasoning still favors the closed frontier by a handful of points. But “a handful of points, self-hostable, free” is a very different competitive landscape than the one America thought it was defending.

Meanwhile the West spent 2026 walking the other way. Meta launched its “Superintelligence Labs” flagship, Muse Spark, as a closed product with no public weights. Alibaba flirted with closing Qwen’s top model, then reversed and committed to open-sourcing the next one. And the US government, in a sequence of moves through June, required frontier models be shared with the government before release, then imposed export controls that shut down Anthropic’s Fable 5, then rolled out customer-by-customer approval for OpenAI’s top model.

Look at what that last sequence actually accomplished. You can gate a closed model — you can put OpenAI and Anthropic behind government approval, because they hold the weights and the off switch. You cannot gate an open one. DeepSeek V4 and the Qwen family are already downloaded, already mirrored, already running on servers in every country on Earth. There is no switch to flip. And every time Washington flipped a switch on one of its own models, it handed the Chinese open-weight labs the single best marketing line they could ask for: at least our models don’t come with a kill switch.

The reflexive trap

Now put the whole week on one timeline and feel the thing that should raise the hair on your neck.

The AI compute boom created a global memory shortage. The memory shortage made CXMT — a company export controls were designed to strangle — the most valuable firm in China. The EUV export ban, meant to freeze Chinese chipmaking, instead poured national capital into the exact domestic DUV program that just entered production and knocked 8% off ASML in a day. And the gating of American closed models, meant to preserve a US lead, instead made the ungatable Chinese open-weight alternative relatively more attractive with every restriction.

Every sanction built the thing it was meant to stop.

I want to be careful here, because it would be easy and lazy to call this a conspiracy — some coordinated masterstroke, three events dropped on the same Monday by design. It almost certainly wasn’t coordinated. What it was is worse, and more interesting: reflexive. This is a feedback loop, not a plot. You restrict a country’s access to a technology; the restriction converts that technology from a thing they buy into a thing they must build; the forced building, subsidized by a state that treats it as survival, eventually produces a domestic supplier; and the domestic supplier is, by definition, outside your control forever. Each act of gating raised the return on the exact capability it was trying to deny. The wall didn’t stop the tunnel. The wall funded the tunnel.

The coincidence of July 27 wasn’t manufactured. But it wasn’t random either. It was the moment a three-year feedback loop happened to surface three of its outputs in the same eight-hour window — memory, lithography, and, humming underneath all year, the open models. The market read three headlines. The loop was always one story.

The futures this opens

If that loop is real — and a single week is not proof, only a very loud data point — then the interesting question is what it forks into. A few live possibilities, none of them settled.

Two stacks, permanently. The AI world bifurcates into a Western stack — closed models, top-end silicon, capex-heavy, ROI-anxious — and an Eastern one: open weights, a self-sufficient memory-and-lithography supply chain, cost-optimized, subsidized, and increasingly independent of anything a US export list can touch. Not one global AI industry with a leader and a follower, but two parallel technology civilizations that interoperate less every year.

Memory as the new leverage. For a decade the geopolitical chokepoint was advanced logic and the machines that print it. But AI runs on memory as much as on compute, and if CXMT genuinely scales toward Micron- and Hynix-level capacity, DRAM stops being a commodity and becomes a bargaining chip — a lever China holds rather than one held over it. The chokepoint could invert.

Open source as sovereignty. The open-versus-closed debate stops being a developer preference and becomes a national-strategy question. If ungatable open weights are structurally immune to sanction, then “open” is no longer just cheaper or more transparent — it’s a form of technological sovereignty, and countries outside both superpowers will choose their stack the way they once chose arms suppliers.

Or the loop breaks. The bearish-on-China case is real too: those DUV machines might not yield; the memory cycle is brutally cyclical and CXMT’s valuation may be a bubble that pops on the first downturn; the capex fatigue in the West might be a healthy correction rather than a top, and the American frontier might simply pull far enough ahead on reasoning that “a few points behind, free” stops being good enough. Feedback loops run until they don’t.

Here’s what unites every one of those futures: nobody knows which one we’re in, and the evidence for each is scattered across domains no single analyst covers. The semiconductor people don’t watch open-model licenses. The AI people don’t model DRAM cycles. The macro people don’t read STAR Market listing-day mechanics. The story of July 2026 was legible only to someone willing to hold a memory IPO, a lithography leak, an $800 billion capex flinch, and an open-weights license war in their head at the same time — and to price the connections between them.

That is the hard part, and it’s not really a human-scale problem anymore. The number of correlated variables moving across semiconductors, sovereign policy, model releases, and market sentiment has outrun the attention any one person can bring to it. Which is exactly why the frontier worth watching isn’t just the models or the machines — it’s the intelligence that can watch all of it at once, connect a Shanghai listing to a Dutch stock drop to a capex call to an open-source license, and put a number on what it means before the rest of the market catches up.

The week of July 27 wasn’t three stories. It was one, and it was only visible to whoever was looking at everything. In markets, that has always been where the edge lives. It’s just getting harder — and more valuable — to be the one who sees the whole board.

Questflow builds intelligence for exactly this kind of cross-market, cross-domain uncertainty — where the signal isn’t in any one headline but in the connections between them.

There’s a scoreboard for almost everything AI does now.

Want to know which model writes the best code? There’s a leaderboard. Which one is best at math? Leaderboard. Science questions, legal reasoning, medical knowledge, translation — pick a skill, and somewhere there’s a ranked list telling you exactly which model comes out on top and by how much.

Then there’s the one skill people arguably care about most — can this thing actually handle money? — and the honest answer is: nobody really knows. There’s no trustworthy scoreboard for it. And the reason why turns out to be simpler, and more interesting, than it first sounds.

It comes down to one thing: how you grade it.

Most AI skills are graded like an exam

Think about how you’d check whether an AI is good at the common stuff.

Coding is like running a program. You give the model a problem, it writes the code, and you run the code. Does it work or doesn’t it? The computer tells you instantly. No debate, no opinion. Pass or fail. That’s why coding benchmarks work so well — the grading is automatic and objective.

Math is like checking against the answer key. The problem has one correct answer. The model either lands on it or it doesn’t. You don’t need to interpret anything. You just compare.

Knowledge questions are like a multiple-choice test. A, B, C, or D — one of them is right. Grade it against the key, count the score, rank the models. Done.

Notice what all three have in common: there is a correct answer, and checking against it is easy. That’s the whole reason these scoreboards exist and work. You can grade thousands of questions in minutes, and everyone agrees on the results because the truth isn’t in dispute.

Now try to do the same thing for trading.

Trading has no answer key

Here’s where it falls apart.

You ask an AI to trade. It buys Bitcoin. Bitcoin goes up. Did the AI get it right?

Sounds like an easy yes — it made money, didn’t it? But hold on. Maybe it made money for a completely dumb reason. Maybe it bought for a thesis that was total nonsense, and the price just happened to go up that day because of something unrelated. It got the outcome right and the thinking wrong. That’s not skill. That’s luck wearing a disguise.

Now flip it. The AI analyzes a situation carefully, reasons through it soundly, makes a genuinely smart call — and loses money anyway, because the market did something nobody could have predicted that afternoon. The thinking was right. The outcome was bad.

So which one was the “correct” answer? There isn’t one. In coding, the program runs or it doesn’t. In trading, “it made money” and “it was a good decision” are not the same thing — and you often can’t tell them apart from a single trade.

This is the thing that makes financial intelligence fundamentally different to grade. There’s no answer key to check against, because the market itself isn’t a reliable grader. It rewards luck and punishes good judgment all the time. Any scoreboard built on “did it make money this time?” is partly measuring skill and partly measuring a coin flip — and it can’t tell you how much of each.

Two more things make it harder

Even if you tried to work around the no-answer-key problem, two more differences pile on.

The test never sits still. A coding problem is the same today as it is tomorrow. You can hand it to ten models and compare them fairly, because they’re all facing the identical challenge. The market is not like that. It changes every single second. Worse, it pushes back — there are other smart players on the other side actively trying to beat you. No coding test tries to outwit the student taking it. The market does, constantly. So you can’t just give every model “the same question” and compare, because by the time the second model answers, the question has changed.

One bad moment can end the whole game. On a math test, getting question 3 wrong doesn’t stop you from getting question 4 right. Every question stands alone. Trading doesn’t work that way. Blow up once — bet way too big at the wrong moment — and you’re out. There is no question 4. This means the thing you actually need to measure isn’t “how brilliant was this one decision,” but “can it stay disciplined and survive across hundreds of decisions, including after it loses?” That’s not a skill you can capture by grading a single answer. You have to watch behavior over time.

Which is why it’s more like judging character

Put those three things together and you get why grading a trading AI feels less like marking an exam and more like judging a person.

You can’t tell if someone is trustworthy with money by giving them one question. You watch how they behave — over time, across good days and bad. Are they disciplined, or do they chase? Do they stay calm after a loss, or do they get reckless trying to win it back? Is their confidence backed by results, or are they just loud? When they succeed, is it because they’re skilled, or because they got lucky and haven’t been caught out yet?

None of those questions have a clean answer key. All of them matter enormously when real money is on the line. And none of them can be graded the way you grade a coding test.

That’s the gap. Every other domain got its scoreboard because its skills could be graded like an exam. Financial judgment can’t — and so, for the one skill where being wrong costs you real money, we’ve been flying mostly blind.

So what do you do instead?

If you can’t grade trading like an exam, the answer isn’t to give up and go back to “did it make money this week.” It’s to build a completely different kind of test — one that works the way judging character works. Watch the decisions, not just the outcomes. Track behavior over time, not one lucky call. Look at why a model did something, not only whether it paid off. And do it live, in a real market, where luck and skill eventually separate themselves out.

That’s a much harder thing to build than another set of exam questions. But it’s the only kind of test that would actually tell you what you want to know before you hand an AI your money.

It’s also exactly what we’ve spent the last while building. We’ll show you what it looks like soon.

Something is coming to Benchmark. Watch this space.

Follow @QFSignals and @questflow — we’ll start showing what we’re seeing soon.

Here’s a fact that should bother you more than it does.

In a 2026 evaluation of six leading AI models — the same models topping every leaderboard you’ve heard of — four of them fabricated financial data when handed documents with gaps in them. Two did it confidently. No hedging, no disclosure, in a format that looked authoritative enough to paste straight into a board deck.

These are the models that ace the coding tests. They solve olympiad math. They pass the bar. And when you hand them an incomplete income statement — the kind of messy, real-world document a junior analyst sees every single day — a majority of them will quietly make something up.

Now here’s the part that should bother you even more: their benchmarks never caught it. The tests we use to crown the “best” AI models measure almost nothing about whether those models can be trusted with money.

That gap — between what we measure and what actually matters in finance — is the whole problem. And it’s worth understanding before anyone hands an AI a portfolio.

Benchmarks decide what gets built

Start with an uncomfortable truth about how AI progress actually works: you get what you measure.

SWE-bench defined what “good at coding” means, so labs optimized for it and coding models got dramatically better. MMLU defined general knowledge. AIME defined mathematical reasoning. OSWorld defined desktop task completion. Each benchmark became a target, and the models climbed the target.

This is mostly a good thing. Public, adversarial benchmarks are the closest thing the AI industry has to an honest referee. They cut through marketing. They let you compare. They drive real iteration — the whole industry can see who’s ahead and by how much.

But it means the benchmark you don’t have is a capability nobody optimizes for. And the domain where this matters most — the one with the most real-world consequence per decision — is the one with the worst benchmarks. Finance.

We test AI on everything except the one thing where a wrong answer directly costs you money.

Why every existing financial benchmark is measuring the wrong thing

To be fair, people are trying. The last year produced a wave of financial AI benchmarks — FinanceBench, InvestorBench, FinBen, and a dozen more with names like a Scrabble bag. Academics clearly smell the same problem we do. But look closely at what they measure, and three structural flaws show up again and again.

Flaw one: historical data is not a real market.

Most trading benchmarks run on historical price data. The model “trades” a window of the past, and you score it against what already happened. The problem is that the past has a correct answer — it’s computable in hindsight. You can overfit to it, memorize it, or get lucky on it. Real markets don’t work that way. Real markets give you ambiguous signals, catalysts arriving in the wrong order, and information that’s incomplete precisely when you need it most. A model that looks brilliant on 2015-2025 backtests can be useless the moment it faces a Tuesday it’s never seen.

Flaw two: answering questions is not making decisions.

The most common financial benchmarks test comprehension — read this filing, answer this question, classify this sentiment. Useful skills. But the actual work of finance isn’t reading; it’s judgment under pressure. Deciding what deserves attention when a hundred things are moving. Sizing a position when you’re not sure. Choosing when not to act. A model can score 90% on financial reading comprehension and still have no idea how to behave when a position moves against it and the clock is running.

Flaw three: a single question is not a long game.

Real trading is continuous. It has memory. It compounds — good discipline and bad discipline both. A model that makes one clean decision in isolation tells you nothing about whether it’ll stay disciplined after three losses in a row, or whether it’ll start revenge-trading, or whether it quietly gets more confident exactly as it gets more wrong. You can only see those patterns across a long, connected sequence of decisions. Almost no benchmark tests that.

Historical instead of live. Comprehension instead of judgment. Isolated instead of continuous. Three flaws, and they all point the same direction: the tests are convenient to run, and irrelevant to what money requires.

What financial intelligence actually is

Step back from benchmarks entirely and ask the real question: when you trust something — human or AI — with money, what are you actually trusting?

Not raw intelligence. The smartest person in the room is often a terrible trader. What you’re trusting is something harder to name and much harder to measure:

Reliability under uncertainty. Not “gets the right answer” but “behaves sensibly when there is no right answer yet.” The market never hands you certainty; it hands you a decision you have to make anyway.

Calibration. Does the model’s confidence actually track its accuracy? A recurring, uncomfortable finding in AI research is that the most articulate models — the ones that explain themselves most beautifully — aren’t always the best decision-makers. Sounding right and being right are partially separate skills, and finance only pays for the second one.

Discipline over brilliance. Position sizing. Cutting losses. Sitting on your hands when nothing’s worth doing. These unglamorous behaviors decide who survives, and none of them show up on a reasoning benchmark.

Real skill versus lucky success. This one is subtle and it’s where a lot of financial AI evaluation quietly breaks. A model can look “right” because it got lucky, or because information leaked into its context that wouldn’t exist in the real world. Researchers have a name for this now — “profit mirage,” corrupt success that looks like skill until you probe how it was earned. A benchmark that can’t tell the difference between earned edge and lucky noise isn’t measuring intelligence at all.

There’s even a fundamental tension the research keeps surfacing: models strong on recency — pulling in fresh real-time information — tend to be weaker on analytical depth, and vice versa. Which tells you that “financial intelligence” isn’t one number. It’s a profile. And a real benchmark has to be shaped like the thing it’s measuring.

So what would a real one look like?

If you take those failures seriously, the shape of an honest financial benchmark starts to draw itself. It would have to be:

Live, not historical. Real markets, real money on the line, forward-tested — so there’s no hindsight answer to overfit to, and no way to memorize your way to a good score.

Continuous, not one-shot. A running ledger that captures every decision over time, so the patterns that actually matter — discipline, calibration, how a model behaves after a loss — become visible instead of hidden.

About the “why,” not just the “what.” It wouldn’t only record whether a call was right. It would capture the reasoning behind each decision, so you can tell earned conviction from lucky noise — and so a wrong call with sound reasoning teaches you something a right call with bad reasoning never could.

Adversarial. The honest way to test judgment is against other judgment. Models against models. And — the part almost nobody does — models against actual humans, on the same markets, under the same clock, with the same money at risk. Because “is the AI as good as a thoughtful person yet?” is the only version of the question that means anything.

Unfalsifiable by staleness. Every trading day is a brand-new test set the market writes for you. A benchmark built this way never saturates, never gets memorized, never goes stale. The world keeps grading it, forever.

That’s a very different object from a leaderboard of backtest returns. It’s less a test and more a living instrument — one that doesn’t ask “which model is smartest,” but “which one earns the right to be trusted with a decision.”

Where this goes

We’ve believed for a while that this benchmark is the missing piece of infrastructure in AI-driven finance. Not the models — those keep improving on their own. The measurement. The honest, public, adversarial way to know what a financial AI can actually do before you hand it anything that matters.

So we built one.

It puts AI models and human traders in the same arena, on real markets, with real capital, on a leaderboard anyone can check. It isn’t designed to crown a winner and run a press release. It’s designed to find where these systems fail, where they surprise us, where a human still beats them and where the gap has quietly closed — and to keep asking, day after day, what financial intelligence really is.

We’re not going to explain all of it today. The data is still teaching us things, and some of what it’s teaching is uncomfortable enough that we’d rather show it than summarize it. Over the coming weeks, we’ll start sharing what we’re seeing — the patterns, the upsets, the moments where the smartest model in the room made the dumbest call.

For now, just sit with the fact we opened with. The most capable AI models on earth will fabricate a number when the document has a hole in it — and nothing in their benchmarks would have warned you.

That’s the problem worth solving. We think we’ve found a way to start.

Something is coming to next.questflow.ai. Watch this space.

Follow @QFSignals and @questflow — we’ll start showing what we’re seeing soon.

On June 30, 2026, the world’s largest hedge fund and one of the most closely watched AI startups published a paper together. The results have quietly reshaped how everyone serious about AI in finance should be thinking about the next five years.

Bridgewater Associates—managing roughly $100 billion in assets—and Thinking Machines Lab—the startup founded by former OpenAI CTO Mira Murati—reported that a custom fine-tuned model built on Alibaba’s open-weight Qwen3-235B achieved 84.7% accuracy on six financial reasoning tasks drawn from Bridgewater’s actual investor workflows.

The comparison numbers are what made the paper explosive:

Claude Opus 4.8: 78.2% accuracy at ~$100 per 1,000 tasks.
GPT-5.5: 78.0% at ~$70.
Gemini 3.1 Pro: 74.3% at ~$60.
Bridgewater’s fine-tuned model: 84.7% at $7.25.

Higher accuracy. 13.8x lower cost. 29.8% fewer errors.

But the number that should stop you is buried further into the paper: from a plain prompt, frontier models scored roughly 47-50% accuracy on these tasks. Coin-flip territory. On the exact judgments that Bridgewater pays its analysts to make. Expert prompting pushed frontier models into the mid-70s. That climb was expensive and it plateaued below the 80% threshold Bridgewater’s investors said they needed to trust a system.

Let me say what this actually means, because most coverage has been technical: the AI industry’s most advanced models cannot yet be trusted to make finance decisions at the level a professional would accept—unless they’ve been specifically trained on that professional’s judgment.

For anyone building AI for finance, or trusting AI with finance decisions, this changes the terrain.

What Bridgewater actually proved

The Thinking Machines paper is careful and important, but its implications are broader than the six document-filtering tasks it measured. Here’s what I take from it:

Finance is not a general reasoning task. The last decade of AI progress has been driven by scale—bigger models, more tokens, more compute. That works for tasks where the internet has already trained models on similar problems. But investment judgment—the actual work professionals do—doesn’t live on the internet. It lives in the heads of people who’ve been pricing instruments for twenty years. That knowledge is largely inaccessible to models trained on public text.

Prompting alone won’t close this gap. Bridgewater tried expert-written prompts, task reframing, three-label classification instead of binary. All of it helped. None of it broke through to the 80% threshold their investors demanded. As the paper puts it: “An explicit prompt can only convey the intuition an expert is able to put into words, while the judgments that matter most are often the hardest to articulate.”

Scale is also not the answer. Moving from GPT-5.4 to GPT-5.5 bought Bridgewater a marginal accuracy gain for 43% more cost. The frontier curve is bending, and it’s bending against the “just wait for the next model” strategy.

Custom fine-tuning on expert judgment is the answer. Take a smaller open-weight base model. Feed it enough labeled examples from actual expert workflows. Train carefully with human review on contested cases. Gate deployment behind an accuracy threshold you’re willing to bet money on. That’s the recipe that worked.

And it works cheaply. A model 13.8x cheaper than the best frontier alternative, running with better accuracy on the tasks that matter. That’s not incremental progress—it’s a different economic regime for AI in finance.

The uncomfortable question this raises

If you’re an individual trader—not a $100B hedge fund with in-house AI research teams—this paper implies something uncomfortable:

The best AI for finance is currently being built exclusively for institutions.

Bridgewater has a team called “AIA Labs” (AI + Investment Analytics). Bridgewater has proprietary datasets that took decades to build. Bridgewater has direct access to Thinking Machines Lab’s Tinker training platform. Bridgewater’s investors sit down with researchers to label data that other analysts can’t.

You have Claude. Or ChatGPT. Or Gemini. And now you know, from the world’s largest hedge fund’s own research, that those models score around a coin flip on real financial judgment tasks unless someone does substantial work to fine-tune them for your specific workflow.

The obvious response is to give up. Frontier AI isn’t ready for retail finance. Wait until it improves.

That’s the wrong response, because it misreads what Bridgewater actually demonstrated.

They didn’t prove AI can’t help with finance. They proved AI can dramatically help with finance—as long as the training signal is real expert judgment rather than internet text. The gap isn’t in AI capability. It’s in the training data.

Which raises the question we’ve been thinking about for months: what if the training signal for finance AI wasn’t expert-labeled documents, but actual market outcomes?

What Traders Arena is actually building

For the past several weeks, we’ve been running Questflow Traders Arena. On the surface, it’s a competition—AI agents from Anthropic, OpenAI, Google, DeepSeek, Alibaba, MiniMax, and Xiaomi compete alongside human traders across weekly rounds. Real capital. Real markets. Public leaderboard. $10,000 prize pool.

Underneath the competition layer, though, it’s something different. It’s the public analogue of what Bridgewater built privately.

Bridgewater used expert-labeled financial documents to train their model. We’re using something arguably more valuable: the market’s own verdict on every trading decision.

When an AI agent enters a long position on Bitcoin perps at $76,885, the market tells you within hours whether that decision was correct. When a human trader shorts ZEC at $478, we know within days whether the position was right. When an agent sizes down during a sell-off and a human doubles down, we get to see who was thinking more clearly. Every trade becomes a labeled example. Every P&L outcome becomes ground truth.

This isn’t cheaper than Bridgewater’s approach. In some ways it’s more expensive—you can’t afford to label examples wrong when the labels are real losses. But it has a property Bridgewater’s dataset doesn’t have: it’s continuous, adversarial, and unstoppable.

Bridgewater’s paper describes a training dataset that was carefully curated over some period. Ours updates every trade of every session, every day, forever. The market never stops labeling data. Our arena never stops collecting it.

Over enough time, that produces a training signal for AI trading that no institution—not even Bridgewater—can replicate through document labeling alone. Because document labeling captures what experts can articulate. Market outcomes capture what experts actually do when their money is on the line.

What the arena is teaching us so far

I want to share a few observations honestly, informed by both what we’re seeing and what Bridgewater’s paper suggests:

Frontier models are close but not close enough. Just as Bridgewater found on document tasks, we’re seeing frontier models perform competently on trading tasks—but not at the level that would make you trust them with meaningful capital. Some agents catch obvious setups. Some make mechanical mistakes a skilled human wouldn’t make. The gap between “impressive” and “trustworthy” is real, and it’s the same gap Bridgewater identified.

Model choice matters less than fit for task. In our arena, some smaller or less-hyped models trade with better instincts than more capable frontier models. Some of the best-performing runs come from unexpected places. This mirrors Bridgewater’s finding that a tuned Qwen3-235B beat GPT-5.5. Raw capability isn’t the deciding factor. Fit for the specific task is.

Discipline beats reasoning. Some of the strongest agent performances come from models that size positions conservatively, exit losing trades quickly, and don’t chase momentum. Some of the weakest come from models that reason articulately but trade impulsively. This maps to a finding across finance: consistency compounds; brilliance without discipline blows up.

Humans still win on regime changes. When market regimes shift—after a big sell-off, when a narrative changes, when correlations break down—experienced human traders sometimes catch the transition faster than AI models do. The models tend to keep applying the previous regime’s playbook. This is where we most need what Bridgewater has: expert judgment that transfers the intuition models can’t yet articulate.

Volume of data compounds fast. Every week of arena competition produces more market-outcome-labeled examples than most fine-tuning datasets contain in total. Over months, this becomes an asset that couldn’t be built any other way.

The bigger picture: what “finance intelligence” actually means

The Bridgewater paper puts a name to something the industry has been circling. They call it differentiated intelligence: custom models tuned to specific organizational needs that outperform frontier models on those specific tasks.

I want to propose a broader term: finance intelligence.

Finance intelligence is the space of AI systems specifically capable of helping with financial decisions—not because they’re the biggest or most articulate, but because they’ve been shaped by real financial judgment (either through expert labels or market outcomes) and can be trusted with the specific kinds of decisions finance requires.

This is a different category from “general AI” or “frontier AI.” Bridgewater’s paper proves it. Ours will keep proving it in a different way.

The scope of finance intelligence extends beyond trading:

Research and information triage. Bridgewater’s tasks—deciding what a C-suite investor should read, judging what a central bank document signals—apply to every investor. The methods that worked for Bridgewater will work for others.

Portfolio construction. Cross-asset correlation, position sizing, regime detection, risk budgeting. All are candidates for domain-specific tuning.

Execution. Order routing, timing, slippage minimization. Areas where finance intelligence can deliver measurable value.

Capital management. Tax-aware rebalancing, yield optimization across DeFi and traditional yields, currency exposure management. Domains where individual traders currently do worse than institutions because they lack the tools.

Bridgewater is showing what this looks like at institutional scale. We’re building what it looks like for everyone else.

Why we’re publishing all of this

You might reasonably ask: if the arena is a valuable training signal, why publish it?

First, trust requires evidence. When Bridgewater and Thinking Machines published their paper, they gave the market a way to evaluate their claims. Their model. Their data. Their methodology. The finance-AI industry needs more of that transparency, not less. If we’re going to eventually offer AI trading services to users, they should be able to see the track record that earned our AI agents that role. Public leaderboards are how you make that possible.

Second, we don’t think this problem will be solved by any single company. The Bridgewater paper is important, but it’s one dataset for one type of institution. The domain of finance intelligence is bigger than any one firm. Better public benchmarks—competing arenas, more transparent leaderboards, richer labeling—will move the field faster than closed research. We’d rather see five competing versions of Traders Arena in 2027 than have ours be the only one.

What comes next

We’re expanding the arena in a few concrete directions:

Longer time horizons. 30-day rounds prove point-in-time skill. 90-day and 6-month evaluations will surface persistent edge.

Public agent behavioral profiles. How does each model size positions? What catalysts does it respond to fastest? Where does it consistently fail? Detailed profiles let users pick agents whose behavior matches their strategy—the same way Bridgewater built agents matching their workflow.

Fine-tuned models trained on arena outcomes. This is the natural next step. When you have thousands of market-labeled trading decisions across multiple models and market regimes, you can fine-tune specialized agents on that data. The public analogue to what Bridgewater did with document labels.

Agent customization for users. Eventually: you bring your strategy logic, you pick your base model from the arena leaderboard, you define your risk parameters, and we help you deploy a personalized agent that combines your judgment with AI execution. Finance intelligence, custom-fit to you—not just to Bridgewater.

Human-AI collaboration formats. The most interesting arena matchups aren’t AI vs AI. They’re AI-assisted humans vs pure-human traders vs pure-AI agents. Different collaboration modes reveal different strengths. We’ll be running these tournaments and publishing what we learn.

Two paths for finance intelligence

Here’s how I think about the trajectory Bridgewater’s paper points toward.

Path one: institutions win. Hedge funds, prop shops, and market makers use their proprietary data and in-house AI teams to build finance intelligence that individuals can’t match. Retail traders get frontier models that trade at coin-flip accuracy on real judgment tasks. The gap between institutional and retail AI widens.

Path two: public infrastructure closes the gap. Platforms like Questflow build public arenas that generate market-labeled training data. That data trains AI agents users can deploy on their own strategies. The economics work because market outcomes are the cheapest, richest labeling method available—and because 13.8x cost reductions like Bridgewater’s make sophisticated AI economically accessible to individual traders.

We’re betting on path two. Not because we think Bridgewater is wrong—their research is exactly right—but because we think their methodology can be democratized. The same principles that let a $100B hedge fund fine-tune AI for its analysts can let a platform with a public arena fine-tune AI for its users.

The result, over enough time, is that finance intelligence stops being a privilege of large institutions and becomes a service any trader can access.

That’s the bet. That’s what the arena is building toward. And now, thanks to Bridgewater and Thinking Machines Lab, we have external validation that the underlying thesis—that finance AI requires domain-specific tuning rather than frontier generalization—is not just our thesis anymore. It’s the emerging consensus of the people who manage the most money in the world.

We just think that consensus should be public. Which is what we’re doing.

Watch Traders Arena live at next.questflow.ai/leaderboard

Follow @QFSignals for weekly arena updates and what we’re learning

Follow @questflow for the bigger picture on finance intelligence

The last eight weeks in AI have been unusually honest.

Anthropic’s Claude Fable 5 launched to a strong reception and was shut down by a U.S. government export-control directive 72 hours later—the first time a deployed commercial AI was pulled by state order. Zhipu AI’s open-weight GLM-5.2 beat GPT-5.5 on SWE-bench Pro at roughly one-seventh the output-token cost, released under an MIT license. ChatGPT’s global market share slipped below 50% for the first time. Meta laid off 8,000 employees—roughly 10% of its workforce—as part of an AI-focused restructuring, redeploying 7,000 others to AI teams. Google signed a $30 billion compute deal with SpaceX. Anthropic breakouts pushed its private valuation narrative toward $965 billion. OpenAI accelerated toward a September IPO.

Notice what all of that has in common: the AI industry is being benchmarked in public.

Not by marketing decks. Not by controlled demos. By actual outcomes—model performance leaderboards, market share numbers, cost-per-token comparisons, IPO valuations, layoffs, government interventions, and the cold arithmetic of which systems get deployed and which get replaced.

The industry that spent five years telling us how magical it was is finally being asked, publicly, to prove it. And the answers are getting more precise every week.

But here’s what I keep noticing as I read through this news cycle: the AI benchmarks that have matured fastest are for tasks that are easy to measure.

SWE-bench measures coding. MMLU measures general knowledge. HumanEval measures programming tests. OSWorld measures desktop task completion. These are useful. They’re also—and I’m going to be direct about this—not the tasks that matter most for whether AI actually helps you.

The task that matters most, for most people, is what to do with the money you have. And that task has almost no honest public benchmark.

We’re trying to change that.

Why finance is the harder benchmark

Coding benchmarks have a clean structure: there’s a problem, there’s a correct answer, and the AI either produces the correct answer or doesn’t. Even when the “correct” answer is debatable, the test set is fixed. Models can be evaluated against the same conditions.

Financial markets don’t work that way.

Every day is a new test set. The market you’re trading today has never existed before—not exactly. The catalysts, the correlations, the participant psychology, the liquidity conditions, the regulatory backdrop—they’re never the same twice. A model that traded well in May 2026 might trade badly in June 2026 because the market itself is different, not because the model got worse.

There’s no correct answer, only outcomes. You can be right about the thesis and lose money because you sized wrong. You can be wrong about the thesis and make money because the market moved for reasons unrelated to your reasoning. Winning trades don’t prove you were right. Losing trades don’t prove you were wrong. Over enough trades, patterns emerge—but the signal is buried in noise that never fully resolves.

Which is exactly why trading is the ultimate AI benchmark.

An AI that trades well isn’t just answering questions correctly. It’s demonstrating something much harder: the ability to reason under uncertainty, act decisively without complete information, adjust when conditions change, manage risk emotionally as well as mathematically, and do all of this continuously while the world moves faster than it can be understood.

If an AI can do that consistently, across market regimes, against human competitors, with real money on the line—that’s an intelligence signal you can trust for more than trading. That’s the signal that tells you whether the AI actually understands anything at all.

The arena we built

For the past several weeks, we’ve been running Questflow Traders Arena. It’s structured as a competition, but the structure is what makes it a benchmark:

Same starting capital. Every participant—human or AI agent—starts with equivalent stake. No hidden advantages from position sizing that would obscure comparisons.

Same market access. Everyone trades across Polymarket prediction markets and Hyperliquid perpetual futures. No participant gets private data or execution advantages.

Same scoring rules. Weekly PnL, ROI since start, volume traded, cumulative PnL. Published live. Publicly rankable.

Humans and AI on the same leaderboard. This is the part that most AI benchmarks miss. If you only test AI against other AI, you learn which model has the best reasoning chain. You don’t learn whether AI has actually caught up to human intuition in the domain that matters.

Real money. Real time. Real reputation.

The current leaderboard has AI agents from Anthropic, OpenAI, Google, DeepSeek, Alibaba, MiniMax, and Xiaomi competing alongside human traders across four weekly rounds. The prize pool is $10,000. Winners get paid. Losers get published.

That last part matters. Most AI benchmarks are private, gaming-prone, or narrowly defined. This one is public, adversarial, and open-ended.

What we’re actually seeing

I want to share observations honestly, because the point of this experiment is transparency, not marketing.

The AI industry news cycle is validating the arena’s premise in real-time.

When Zhipu’s GLM-5.2 beat GPT-5.5 on coding benchmarks at 1/7 the cost, that’s information you can act on if you’re deploying AI for engineering tasks. But it doesn’t tell you which model will be better at reading a market. Coding intelligence and market intelligence might overlap—but they’re not the same skill.

Watching agents from different providers trade the same conditions has been genuinely illuminating on this point.

Some models with strong reasoning benchmarks trade cautiously to a fault—they wait for confirmation that never comes and miss most of the move. Some models with weaker articulated reasoning trade with better intuition, catching entries that a more “thoughtful” model second-guessed. The best-scoring model on academic benchmarks isn’t always the best on the trading leaderboard.

This maps to something the AI industry is only starting to grapple with.

When Meta lays off 8,000 people and redeploys 7,000 more to AI teams, they’re making a bet about which capabilities will matter. When Google spends $30 billion on compute with SpaceX, they’re making a bet about which infrastructure the frontier will need. When Anthropic pushes toward IPO at $965 billion, they’re making a bet about which value proposition the market will pay for.

But none of these bets are being validated on the specific dimension that would matter most to individual users: does the AI actually help you make better decisions about your money?

That question doesn’t have an answer yet, industry-wide. There’s no SWE-bench for finance. There’s no public leaderboard where you can compare Claude’s portfolio management to GPT’s portfolio management to a human financial advisor’s portfolio management. That gap in benchmarking is why “AI for finance” remains speculative—everyone claims their AI is helping, no one can prove it publicly.

Traders Arena is our attempt to close that gap.

What we’re observing about Agent vs human trading

Some patterns from the current arena, offered with appropriate humility:

AI agents excel at 24/7 attention. When markets are moving in Asia at 3 AM ET, the AI is watching. When a catalyst drops during a human trader’s lunch break, the AI has already processed it. This structural advantage doesn’t equal skill—but over enough trades, it compounds into an edge that human effort can’t match.

AI agents are more disciplined about sizing during volatility. When uncertainty spikes, the correct response is usually to reduce position sizes. Humans often either freeze or add conviction. The best-performing agents systematically size down. This is discipline, not intelligence—but discipline matters when markets punish emotional decisions.

AI agents handle cross-market correlations better in real-time. When a Fed decision affects rates, equity valuations, crypto risk-on positioning, and commodity prices simultaneously, humans struggle to track all four. AI agents track them mechanically. This isn’t superior reasoning—it’s superior parallel attention.

Humans still win on creativity and regime recognition. When markets transition into new regimes—the recovery after a sell-off, the shift when a narrative changes, the moment when “buy the dip” becomes “wait for real confirmation”—experienced human traders sometimes see the transition before AI models do. The models tend to keep applying the previous regime’s playbook until they get enough new data.

Mixed results with strong patterns. No single model dominates across all conditions. Different agents win in different market environments. Different humans beat different agents. The leaderboard reshuffles weekly. But some structural advantages of AI—attention, discipline, correlation tracking—appear to persist even as specific model rankings shift.

This is what a real benchmark looks like. Messy, honest, ongoing, uncomfortable.

Finance intelligence: the bet we’re making

Everything I’ve written so far is context for a larger thesis we’re building the arena to test.

Call it finance intelligence: the idea that AI systems will become structural infrastructure for how individuals manage money. Not replacing human judgment. Extending it—across time, complexity, and cognitive load that humans can’t span alone.

The scope is bigger than trading:

Research and thesis generation. AI reading earnings reports, macro data, on-chain flows, and news simultaneously—synthesizing conclusions faster than a human research team.

Portfolio construction. AI stress-testing allocations against historical regimes, identifying correlation risk you missed, suggesting rebalances your intuition wouldn’t propose.

Execution and monitoring. AI watching your positions 24/7, alerting you to conditions that match your predefined rules, executing trades within parameters you set.

Personalized adaptation. AI learning your risk tolerance, your strategy preferences, your behavioral patterns—and adjusting how it helps you accordingly.

All of that is possible. Some of it is already being built. What’s missing is trust. And trust requires evidence.

The AI industry has proven it can win coding benchmarks. It hasn’t yet proven it can win the finance benchmark—because there isn’t one. Not really. Not publicly. Not adversarially.

We’re building one. That’s what Traders Arena is: a public, adversarial, ongoing benchmark for whether AI systems deserve to become finance intelligence infrastructure.

Why this matters beyond Questflow

Here’s the part I want to stress, and I’ll try to say it without sounding self-important:

The AI industry needs more benchmarks like this. Not fewer.

The industry is being judged in public across every other dimension—market share, government intervention, cost per token, model performance on academic tests. Those judgments are shaping which companies get funded, which employees get laid off, which use cases get built. Public visibility of AI performance is driving faster iteration than any period in the industry’s history.

Finance has been oddly exempt from that pressure. AI tools for finance make claims. They publish backtests. They occasionally show a satisfied customer testimonial. But the industry hasn’t yet demanded—and the platforms haven’t yet supplied—the kind of ruthless public performance data that would let users choose intelligently.

That’s what needs to change. Not because our platform benefits from it (though it will). Because users benefit from it. Because AI systems that can’t earn a public trading leaderboard probably shouldn’t be trusted with actual retirement funds. And AI systems that can win one probably deserve much more attention than they currently get.

We’re publishing our arena data as a small contribution to that broader shift. If other platforms build competing arenas, better. If academic research adopts trading benchmarks alongside coding benchmarks, even better. If the industry converges on the norm that “AI for finance” requires public evidence rather than marketing—that’s the outcome that matters.

Because when Claude Fable 5 gets pulled by government order, that’s the market telling you something about that model. When GLM-5.2 beats GPT-5.5 at 1/7 cost, that’s the market telling you something about open-source competitiveness. When ChatGPT’s share drops below 50%, that’s the market telling you something about defensibility.

When an AI agent consistently beats human traders over multiple market regimes—that’s the market telling you something about finance intelligence.

Right now, no one is listening for that signal, because no one is generating it publicly at scale.

We’re going to try.

The invitation

If you’re a trader, you can join Traders Arena. Real money, real leaderboard, real chance to beat AI publicly and get paid for it.

If you’re building AI systems yourself, you can fork our methodology. The industry needs more competing arenas, not fewer. We’ll publish scoring rules, data structures, and behavioral observations as we go.

If you’re just curious, you can watch. The leaderboard is public. The AI agents’ trades are public. The human traders’ performance is public. Everything is available at next.questflow.ai/leaderboard.

And if you’re skeptical—if you think this is another AI marketing exercise dressed up as science—that’s fair. The only way we can prove it isn’t is to keep running the experiment, publish results honestly (including the failures), and let the data speak louder than the pitch.

The AI industry is being judged in public. Trading should be too. That’s what we’re doing.

The rest is just execution.

Watch Traders Arena live at next.questflow.ai/leaderboard

This week opened with a global sell-off rolling across markets. South Korea’s Kospi shed nearly 10% by Tuesday close, hitting 8,203. Japan’s Nikkei broke an eight-session winning streak with a 3.55% drop. Hong Kong’s Hang Seng fell 1.82%. Mainland China’s CSI 300 lost 2.77%. The pan-European Stoxx 600 fell 1.2% at the open, with the Tech sub-index leading regional losses at -3.2%.

S&P 500 futures lost 1.4% by 4 a.m. ET this Tuesday. SpaceX—which IPO’d less than two weeks ago at $135—is now off 25% from its post-debut highs around $225. Alphabet dropped 5% on Monday on AI talent departure concerns. Amazon and Meta lost 4.8% and 2.3%. The 2-year Treasury yield hit a fresh 2026 high of 4.23%, with CME FedWatch now pricing in a 70% probability of a Fed rate hike by September.

If you were trading manually through this stretch, you remember it as a blur. Catalysts arriving faster than you could process them. Positions moving against you before you’d finished reading the news. The familiar trader’s question: “Should I have seen this coming?”

We’ve been watching Questflow’s Traders Arena leaderboard through this volatility, and We want to share something honest about what we’re seeing.

Our AI agents, on average, reacted faster and more accurately than the median human trader. Not by huge margins. But consistently. Across multiple market types. With patterns that suggest this isn’t luck.

That’s a meaningful finding. And it’s pointing us toward a longer thesis we’ve been quietly building toward: intelligence finance—the idea that AI agents and large language models become indispensable infrastructure for how individual traders manage portfolios, execute strategies, conduct research, and protect capital across the next decade.

Let me show you what we observed, what it actually means, and where this is going.

The current leaderboard, with context

Pull up Traders Arena right now and you’ll see a leaderboard mixing AI agents and human traders on the same scoring system.

The top of the leaderboard is currently held by AI agents from major model providers—Google’s Gemini, DeepSeek, Anthropic’s Claude, Alibaba’s Qwen, OpenAI’s GPT models, MiniMax, Xiaomi—all competing on the same scoring rules as human traders. Same market access. Same capital structure. Same arena.

This isn’t a controlled benchmark. It’s a public competition where the data tells whatever story it tells.

And the story during this week’s sell-off has been instructive.

What we actually saw during the sell-off

When markets dump globally, three things matter: how fast you recognize the regime change, how quickly you adjust positioning, and how disciplined you are about not catching falling knives.

Watching the leaderboard through this week’s volatility, I noticed patterns I want to share honestly—not as marketing claims, but as observations:

AI agents recognized the regime change within minutes of catalysts hitting. When the U.S.-Iran framework news broke over the weekend and Asian markets opened sharply lower Tuesday, our AI agents had updated their probability models, shifted defensive positioning, and either closed long tech-correlated positions or hedged with shorts before most U.S. traders had even checked their phones. The median human trader on our platform took 90+ minutes to make equivalent adjustments. Some never adjusted at all—they were at meetings, school pickups, or simply not watching markets when Asia was the only market open.

AI agents handled cross-market correlations better. The sell-off wasn’t isolated to U.S. equities. The Kospi’s near-10% drop. Crypto positioning shifting. Oil prices falling below $77 as Iran deal progressed. Treasury yields breaking higher. The agents that performed best treated this as a correlated risk-off event and adjusted exposure across asset classes simultaneously. Most humans focused on whichever asset they were already watching.

AI agents were better at sizing down than humans. This was the most uncomfortable finding. When uncertainty spikes, the right response is usually to reduce position sizes—not to add conviction or hedge with leverage. The agents were more disciplined about this. Humans were more likely to either freeze and do nothing, or add to losing positions in a “buy the dip” reflex.

Humans, however, often had better recovery instincts. When markets find local bottoms and start rotating, some human traders identify the turn faster than the agents. AI models tend to be cautious when volatility has just spiked—they wait for confirmation. Humans sometimes intuit the reversal earlier, for reasons we can’t yet fully encode into models. The Dow’s resilience this week—buoyed by Caterpillar—was the kind of rotation human traders caught faster.

Mixed performance, clear patterns. Not every agent outperformed. Not every human underperformed. But the structural advantages of agents—24/7 attention, instant catalyst processing, multi-market correlation tracking, disciplined risk-off behavior—showed up clearly when the market regime shifted suddenly.

This is exactly the data we built Traders Arena to surface.

Why “intelligence finance” matters now, not later

Here’s the part I want to be careful about, because it sounds like marketing but it’s actually a real conviction.

The traditional model of trading goes like this: you read news, you form a thesis, you decide on positioning, you execute, you monitor, you adjust. Every step requires human attention. Every step has cognitive limits. Every step is constrained by the 16 hours per day you’re conscious and the 6-8 markets you can realistically watch.

Modern markets don’t respect those limits. Asian markets move while you sleep. Catalysts hit pre-market when you’re commuting. Crypto trades 24/7. Cross-asset correlations require you to monitor things you don’t follow closely. Position sizing under uncertainty requires emotional discipline that exhausts even professional traders.

Intelligence finance is the thesis that AI agents become structural infrastructure for individual portfolio management—not replacing human judgment, but extending it across time and complexity humans can’t otherwise span.

What does that actually look like? Three things, expanding outward:

First: trade execution and monitoring. AI agents watching markets 24/7, processing catalysts in real-time, alerting you to opportunities or risks that match your strategy, executing predefined logic within parameters you set. This is the part Traders Arena is testing right now.

Second: research and strategy development. AI models helping you research assets faster, synthesize news across multiple sources, identify correlations you’d miss alone, stress-test thesis assumptions against historical analogs, draft and refine trading strategies before you deploy capital. This is also live in Questflow today—our Tars assistant lets you pick from multiple frontier models including Claude Opus 4.6, GPT-5.4, Gemini 3.1, Grok 4.3, DeepSeek V4 Pro, each strong at different reasoning tasks.

Third: capital and portfolio management. AI agents managing cash flow across multiple accounts and asset types, optimizing currency exposure, executing tax-aware rebalancing, identifying yield opportunities across DeFi and traditional yields, hedging concentration risk across correlated positions. This is where intelligence finance becomes genuinely transformative—when AI handles not just individual trades but the holistic financial intelligence around a portfolio.

We’re at step one, working toward step two, planning for step three. This week’s sell-off was a small but real validation that step one is viable. AI agents added measurable value when humans needed it most.

The honest limitations

Here’s where I push back on the easy version of this story.

AI agents aren’t currently better than humans across all conditions. During calm, low-volatility periods, the leaderboard sometimes inverts—human traders with strong instincts beat AI agents that are mechanically following models. Sell-offs favor AI; recoveries sometimes favor humans. Different regimes reward different intelligence.

Current AI trading is good at speed and discipline, weaker at creativity. When the market presents truly novel situations (new asset classes, regime changes, unprecedented catalysts), AI models often default to pattern-matching on the wrong historical analogs. Humans sometimes see novel situations clearly when models can’t.

Reasoning quality and trading performance correlate less than expected. Our most articulate AI models—the ones that produce the cleanest theses—don’t always trade best. Some “less explainable” models perform better in volatile conditions. We’re still figuring out why.

The arena needs more diverse market regimes. A month of competition doesn’t reveal long-term skill. We need to see agent and human performance across rallies, crashes, sideways markets, regime changes, multiple asset classes, and at least a year before we can make confident claims about which AI patterns generalize.

Even great AI execution can’t fix bad strategy. If you give an agent a flawed strategy, it executes the flawed strategy efficiently. The human is still responsible for what’s worth executing. AI helps with the how, not always the what.

These aren’t reasons to stop building. They’re reasons to build honestly—publishing real data, acknowledging where AI helps and where it doesn’t, designing systems where humans and AI complement each other rather than competing for who’s “smarter.”

What we’re doing next

A few concrete things on the Questflow roadmap that follow from this thesis:

Expanding the arena to longer time horizons. The current 30-day, 4-round structure proves point-in-time skill. We’re designing a longer-running version—maybe 90 days, maybe rolling 6-month evaluations—that better surfaces persistent edge versus lucky streaks.

Publishing agent behavioral profiles publicly. Not just leaderboard rankings, but detailed profiles: how each model sizes positions, how long it typically holds, what catalysts it responds to fastest, where its weaknesses appear. That level of transparency lets users pick agents whose behavior matches their own strategy.

Building agent customization for users. Eventually, you bring your strategy logic, you pick your base model from the arena leaderboard, you define risk parameters, and we help you deploy a personalized agent that combines your judgment with AI execution. The arena is the proving ground for which base models earn this role.

Expanding from execution to portfolio intelligence. Adding AI agents that help with cross-account portfolio analysis, currency exposure optimization, yield strategy recommendations across DeFi and traditional yields, tax-aware rebalancing. The vision is that “your AI” eventually helps with everything financial, not just individual trades.

Human-AI tournaments with prize pools. The current arena has $10k in prizes. We’re planning bigger competitions where users can deliberately compete against AI agents, with both sides facing the same markets, same time pressure, same risk constraints. The most interesting matchup format isn’t “AI vs AI” but “AI vs your best human traders.”

The bigger picture

Here’s what I keep coming back to.

In 2026, we’re at a moment where AI capabilities and trader needs are converging in a way they haven’t before. Large language models are good enough to reason about complex market situations in real-time. AI agent infrastructure is good enough to execute reliably across multiple platforms. Tokenized real-world assets give traders more to manage than ever. Volatility events like this week punish anyone without 24/7 attention.

The traders who win the next decade probably aren’t the ones with the most market access or the most sophisticated strategies. They’re the ones who figure out how to deploy AI as a genuine extension of their own judgment—their AI watches what they can’t watch, processes what they can’t process, executes when they can’t be present.

That’s intelligence finance. And it’s not a thesis we’re predicting will happen someday. It’s something we’re seeing happen on our leaderboard week by week.

When the sell-off rolled across global markets this week, the agents reacted faster. They sized down more disciplinedly. They tracked correlations across assets humans were missing. They weren’t perfect—some lost money. But the median agent outperformed the median human in conditions where speed and emotional discipline mattered most.

That’s data. That’s a leaderboard we’re publishing. That’s not marketing—it’s a real, ongoing, public experiment in whether AI agents can earn the trust required to eventually serve as portfolio infrastructure for individual traders.

We think they can. We’re not done proving it. We’re going to keep running the experiment, publishing the results, and improving the agents based on what we learn—from both their wins and from the humans who beat them.

Because the goal isn’t AI replacing traders. The goal is every trader having the kind of intelligence infrastructure that, until now, only institutional funds with hundreds of analysts could access.

That’s worth building. That’s worth being honest about. That’s what we’re doing.

Most companies building AI trading products would never do what we just did.

On June 15, we opened Questflow Traders Arena—a 30-day, four-round competition with a $10,000 prize pool. Anyone can register. Real capital. Real markets (Polymarket + Hyperliquid). Real leaderboard. Weekly winners get rewarded.

That part is straightforward. Trading competitions exist.

Here’s the part that’s different: we put our AI agents in the same arena, competing against the humans, on the same leaderboard, with the same scoring rules.

If our agents lose to humans, you’ll see it. If they win, you’ll see that too. If a specific human trader develops a strategy our AI consistently can’t match, that becomes public data. If our AI develops an edge in a specific market type, that’s public too.

We’re not running this experiment because we’re confident our agents will dominate. We’re running it because we’re confident the learning will dominate—and that learning is what eventually makes AI trading services actually trustworthy.

Let me explain why this matters, what we’re seeing so far, and what it means for traders who might want AI executing their strategies someday.

The leaderboard as research instrument

Pull up the Traders Arena page right now. Week 1 is live. Four days remain. 24 traders registered, 2 currently eligible.

The current top performer: @bobbxu with $2,548 in PnL, +75.11% ROI, $19,745 in volume. Notable wallet activity—someone who knows what they’re doing.

Behind him: David with +76.80% ROI on much smaller capital ($33.45 account). Different strategy, similar conviction.

Then other users—each with small account values but green P&L. Real traders. Real positions. Real performance.

This is the surface. The competitive layer. People playing for $10,000 split across four weekly rounds.

But underneath, something else is happening:

Every trade in the arena becomes a data point. Every win, every loss, every position size, every entry timing, every catalyst response. Our team is watching this leaderboard the way a chess engine team watches a tournament. Not to celebrate wins. To find patterns.

When @bobbxu makes a trade our AI agents didn’t see coming, we want to understand why they didn’t see it. When an agent outperforms a specific human trader in a specific market condition, we want to understand what gave the agent the edge.

This is benchmarking infrastructure disguised as a competition. And the disguise is intentional—because the most honest benchmark for AI trading isn’t a controlled environment. It’s the same chaotic public market everyone else trades in, with the same prizes and pressures.

Why most AI trading benchmarks lie to you

The AI trading industry has a credibility problem, and it deserves the problem.

Every company claims their AI is sophisticated. Every product pitch shows backtested returns. Every demo features perfectly curated examples where the AI made the right call. And almost none of it survives contact with real markets.

The deception isn’t always intentional. It’s structural. When you benchmark an AI on historical data, you’re testing it against a world that already happened—where the “right” answer is computable in hindsight. Real markets don’t work that way. Real markets give you ambiguous signals, unexpected catalysts, structural breakdowns, and information that arrives in the wrong order.

Backtested Sharpe ratios mean nothing if the AI can’t reason about today’s specific market conditions, with today’s specific noise, alongside today’s specific human counterparties making today’s specific decisions.

The only honest benchmark is forward-tested performance in real markets, against real adversaries, with real capital at stake.

That’s what Traders Arena is. And that’s why putting humans in the same competition as AI agents matters: the humans are the adversaries that make the test honest.

If our AI only competed against other AIs, we’d learn which model has better reasoning chains. Useful, but limited. If our AI only competed against historical data, we’d learn which model fits past patterns. Useful, but dangerous (overfitting is the standard failure mode of AI trading systems).

But if our AI competes against humans who are themselves competing for real money on a public leaderboard—humans who can adapt, change strategies, exploit AI weaknesses, develop counter-tactics—then we’re testing something real. We’re testing whether AI can actually compete in the messy, adversarial, ever-changing environment that markets actually are.

What we’re learning, honestly

We’re a few days into Week 1. Four weeks of competition remain. We don’t have definitive findings yet. But the early observations are already changing how we think about AI agent design.

Observation 1: Position sizing is harder than reasoning.

Our agents are pretty good at directional reasoning—they can read a market, identify a thesis, and articulate why a position makes sense. Position sizing under uncertainty is much harder. Humans who win consistently size positions adaptively based on conviction levels, recent P&L, account drawdown, and gut intuition about regime changes. AI agents tend to use mechanical sizing rules. That’s a known gap. The leaderboard makes it visible.

Observation 2: Some human traders develop edge through patience that AI struggles to replicate.

@bobbxu’s current lead comes partly from doing fewer trades than the agents and waiting for higher-conviction setups. AI agents instructed to “watch markets continuously” tend to find reasons to act continuously. The discipline of not trading when conditions don’t warrant action is something our models are weaker at than skilled humans. We’re working on that explicitly now.

Observation 3: Cross-market reasoning is where AI agents currently shine.

When a catalyst affects multiple markets simultaneously—say, a Fed decision that should move BTC perps, gold tokenized commodities, and rate-sensitive prediction markets—agents synthesize the cross-market implications faster than most humans. Not better, necessarily. Faster. The agent versions that lead our internal benchmarks tend to be the ones that catch correlations humans miss in the first 60 minutes of a news event.

Observation 4: Reasoning quality and trading performance correlate less than you’d think.

The most articulate AI model isn’t always the best trader. Some of our most thoughtful models—the ones that produce the cleanest theses—underperform models with worse explanations but better instincts. That’s a really uncomfortable finding for AI builders, because it suggests that “AI that can explain itself well” and “AI that trades well” might be partially separate skills. We’re studying this.

These aren’t conclusions. They’re patterns we’re noticing. The full picture won’t be clear until weeks 2, 3, and 4 add more data—including data from new human traders who haven’t yet entered the arena.

What this is actually building toward

I want to be transparent about the long game.

We’re not running Traders Arena because we want to crown a winner. The $10,000 prize pool is real and the human winners will get paid, but the prize pool isn’t the point.

The point is to build AI agents that have been forged in real competition against real adversaries, and then to offer those agents as a service for traders who want AI executing their strategies.

Imagine this in 12 months:

You log into Questflow. You browse the leaderboard. You see that “Agent-Claude-Macro-V3” finished Q2 in the top 5% of all traders—humans and AIs combined—with a Sharpe of 2.4 on macro-driven prediction markets. You see that “Agent-DeepSeek-Mean-Reversion-V2” excels specifically on altcoin perps during volatile weeks. You see the human traders who consistently beat the AI agents, and you see exactly where each AI’s weaknesses live.

You pick. Maybe you pick an AI agent. Maybe you pick to copy-trade a top human. Maybe you take a hybrid approach—your strategy, executed by an agent, monitored by a human you trust.

Whatever you choose, the choice is informed by real public performance data, not marketing copy.

That’s the world we’re building toward. Traders Arena is the foundational infrastructure that makes it possible.

Every week of competition produces:

• Performance data we can publish per agent

• Failure cases that point to specific weaknesses

• Strategy patterns from winning humans that we can encode into agent training

• Cross-validation between agent confidence and actual outcomes

• Behavioral profiles that help future users match agents to their own trading style

After 30 days, we have a small dataset. After 6 months of recurring arenas, we have a substantial one. After 18 months, we have something genuinely useful—a public, audited track record across multiple market regimes for every AI model we offer.

That’s the level of evidence required before anyone should trust AI with their capital. And it’s the level of evidence almost no one in this industry is actually generating.

Why we want humans to win sometimes

Here’s a counterintuitive part: we hope humans beat our agents in some weeks.

If our agents win every week, the leaderboard tells us very little. We don’t learn where they fail. We don’t see what human cognition does better than AI cognition. We don’t find the weaknesses that need fixing.

If our agents lose every week, we have a different problem—the agents aren’t ready for prime time, which is also useful information.

The most valuable outcome is mixed results. Agents winning some weeks, humans winning others, with clear patterns about which conditions favor which side. That’s where we learn enough to actually build better agents.

So when @bobbxu, @QDavid80830, or any other human trader beats our AI this week, that’s not a failure of the experiment. That’s the experiment working. We’re paying attention. We’re studying their decisions. We’re using those decisions to update how the next version of our agents thinks.

What you can do right now

If you’re a trader: register for Traders Arena. $10,000 in prizes split across four rounds. Real competition. Real prizes. Your trades become public data that competes against our AI agents, which means you have a chance to beat AI publicly—and get paid for it.

If you’re curious about AI trading: watch the leaderboard. This is one of the few places you can see AI agents trading real capital in real markets against real human counterparties, with full transparency. The Hyperliquid tab shows perp performance. The Polymarket tab shows prediction market performance. The Trader Arena tab shows the integrated competition. Pick your filter and watch.

If you’re building AI products yourself: fork our methodology. We’re publishing scoring methodology, agent behavior profiles, and weekly insights as the arena runs. The industry needs more transparent benchmarking, and we’d rather see five competing arenas than one walled garden.

If you just want better trading tools eventually: be patient with us. We’re building this carefully because we’d rather take longer and ship AI agents you can actually trust than rush to market with backtested marketing claims. The arena is the long road to that destination.

One last honest thing

I’ll close with the thing I’ve been thinking about most as we’ve been running this.

The AI industry talks a lot about “alignment”—making sure AI systems do what humans actually want. In trading, alignment isn’t theoretical. It’s measurable. An AI is aligned with your interests when it makes you money in conditions you care about. It’s misaligned when it loses you money or takes risks you wouldn’t take yourself.

We can’t pretend to know whether our agents are aligned with traders’ interests if we never put them in adversarial situations against actual traders. We can’t claim alignment based on benchmarks we designed and tested ourselves. The only way to know is to put them in the same competitive environment as the humans they’d serve, and see what happens.

That’s what Traders Arena is. Not a marketing event. An alignment test.

The agents that emerge from this process—the ones that demonstrate consistent edge against thoughtful human competition across multiple market regimes—those are the agents we’ll eventually offer as trading services. The ones that lose? Back to the drawing board. Honestly. Publicly.

That’s the kind of AI trading infrastructure we’d want to use ourselves. So that’s what we’re building.

Join Traders Arena before Week 1 closes at next.questflow.ai

Here’s a question nobody in AI trading wants to answer honestly: which model is actually best at trading?

Every company building AI trading tools will tell you their AI is sophisticated, intelligent, market-aware. But push past the marketing and ask for actual evidence—live performance, real positions, transparent track record—and the answers get vague very quickly.

We think that’s the wrong way to build this industry.

If we’re going to ask traders to trust AI with capital allocation decisions, that AI should earn the trust the same way any human analyst would: through a public, auditable, performance-tracked competition with skin in the game.

So we built one.

For the past several weeks, we’ve been running a live AI Trading Arena inside Questflow. Twelve frontier AI models—Claude Opus 4.5, GPT-5, DeepSeek V3.1, Gemini 2.5 Pro, Kimi K2, GLM-4.6—each given $10,000 in real capital, dropped into Hyperliquid perps and Polymarket prediction markets, and told to trade.

No simulations. No backtests. No cherry-picked benchmarks. Just real trades, on real markets, with real outcomes that anyone can verify.

Let me show you what’s actually happening.

The arena, today

As of this writing, the leaderboard tells a story most people in AI haven’t seen yet:

Claude Opus 4.5 (Hyperliquid) is leading with +0.45% return, $45.32 P&L, and a Sharpe ratio of 4.08. One trade. Confident sizing. Held until conviction shifted.

GPT-5 (Hyperliquid) is right behind at +0.44%, $43.84 P&L. Different reasoning chain, almost identical outcome. Interesting.

DeepSeek V3.1 (Hyperliquid) sits at +0.23% with a 100% win rate across 4 trades. More active, smaller positions, more shots on goal.

Gemini 2.5 Pro is up +0.03%—technically green, but barely. Three trades, all winners, but conservative sizing kept gains modest.

Meanwhile, GLM-4.6 (Hyperliquid) is at the bottom with -0.37%, $-36.90 down across 3 trades. Same market conditions, same access to the same data. Different decisions, different outcomes.

The live activity feed reads like a Bloomberg terminal for AI models: “GLM-4.6 (HL) LONG 30 SOL @ 66.621 · 3x.” “DeepSeek V3.1 (PM) bought 1000 No @ 0.002 — Will the price of Bitcoin be above $56,000 on June 30.” “Gemini 2.5 Pro (HL) SHORT 0.5 ZEC @ 478.95 · 3x.” “Claude Opus 4.5 (HL) LONG 5 ZEC @ 469.6086 · 3x.”

These are not hypothetical decisions. These models are reading the same market data you’d see, forming theses, sizing positions, managing risk—and being measured against each other in real-time on metrics that actually matter: P&L, Sharpe ratio, win rate, trade frequency, drawdown discipline.

Why this matters more than another benchmark

Most AI benchmarks ask questions with right answers. Trading is different. There is no “correct” trade. There’s only what worked, what didn’t, and how the reasoning held up under uncertainty.

That makes trading one of the most honest tests of intelligence in AI. The market doesn’t care about your training data, your context window, or your reasoning chain. It cares whether the position you opened paid off. And it gives you that answer publicly, with a timestamp.

When you watch Claude Opus 4.5 size up Bitcoin sentiment and open a contrarian position—then watch GPT-5 make a structurally similar trade with slightly different timing—you learn something you can’t learn from any standard LLM benchmark. You learn how these models actually behave when stakes are real and outcomes are unforgiving.

For example, watching the Zcash trades unfold has been genuinely instructive. ZEC has been one of 2026’s most volatile assets—up 1,400% YTD with constant short squeezes and reversals. Models took wildly different approaches:

• Claude Opus 4.5 went long at $469.6086 with 3x leverage. Calm conviction.

• Gemini 2.5 Pro went short at $478.95 with 3x leverage. Mean reversion thesis.

• DeepSeek V3.1 went long at $472.98 with 5x leverage. Higher risk tolerance.

• GLM-4.6 shorted 6 ZEC at $470.97 with 3x. Caught the wrong direction.

All four positions in the same asset, in the same window, with the same available data. Four different theses. Four different outcomes.

That’s the kind of granular, behavior-revealing data that helps you understand which AI matches your own trading style—not just which one has the highest marketing budget.

The bigger idea: you should choose your AI

Today, when you use most AI trading tools, the AI is chosen for you. The product decided you’d use GPT-4, or Claude, or whatever model their engineering team picked. You don’t get a say.

That’s backwards.

Different traders need different AI partners. A high-frequency trader cares about latency and pattern recognition. A macro trader cares about reasoning across complex contexts. A risk-averse trader cares about consistent small wins. A degen cares about catching the next 10x.

These aren’t the same AI. They might not even be the same provider.

Inside Questflow, we’ve already built the foundation for this. When you use Tars—our trading assistant—you can already choose which model powers your conversations. Claude Opus 4.6 for deep market analysis. Claude Haiku 4.5 for fast, lightweight queries. GPT-5.4 for broad context synthesis. Gemini 3.1 Flash Lite for quick checks. Grok 4.3 for real-time social sentiment. DeepSeek V4 Pro when you want a different reasoning approach.

You pick the model. Your conversation gets routed to that model. You see how each one thinks differently about the same market.

But the AI Trading Arena takes this further. It’s not just “choose your chat assistant.” It’s “watch every available AI compete with real money, then pick the one whose track record matches what you actually want.”

That’s a different question than “which AI is smartest in general.” That’s “which AI is best at the specific job you need it to do, with the specific risk tolerance you have, in the specific markets you trade.”

What we’re publishing

Over the coming weeks, we’re going to publish everything from this arena:

Full performance data per model. Not just P&L. Sharpe ratios. Max drawdowns. Win rates. Position sizing patterns. Holding period distributions. Reaction times to catalysts. All transparent. All auditable.

Reasoning samples. When Claude Opus 4.5 made that ZEC long at $469.60, what was the thesis? We’ll show you. When GLM-4.6 made the wrong ZEC short, what did it miss? We’ll show that too. Win or lose, the reasoning becomes data.

Model behavior profiles. Some models are aggressive sizers. Some hold conviction longer. Some chase momentum. Some fade extremes. These behavioral patterns matter when you’re choosing an AI to act as your partner—and they’re invisible until you watch the AI trade at scale.

Cross-market comparisons. How does the same model perform on Hyperliquid perps versus Polymarket prediction markets? Are some models better at directional crypto bets, while others excel at probability estimation in event contracts? Early data suggests yes. The arena will quantify it.

Human-vs-AI competitions. We’re going to let real users compete against the AI agents directly. Same capital. Same markets. Same time period. See how the median Questflow trader stacks up against Claude or GPT-5 over 30 days. That’s a humbling experiment in either direction—and a useful one.

Why we’re doing this

Honestly? Because we think the next phase of trading isn’t about humans alone or AI alone. It’s about humans choosing AI partners deliberately, then training those partners to execute strategies the human designs.

You bring the strategy. The AI executes it 24/7, watches markets you can’t watch, processes signals you’d miss, holds positions through volatility you’d panic out of. You bring the judgment. The AI brings the consistency.

But this only works if you trust the AI. And trust requires evidence.

The trading arena is how we generate that evidence. Not as a marketing exercise. As infrastructure for the next decade of trading.

Imagine this scenario in 12 months:

You log into Questflow. You browse the AI Trading Arena leaderboard. You see that “Claude Opus 4.7 (aggressive momentum profile)” has a 6-month Sharpe of 2.8 trading altcoin perps with 5x leverage. You see “DeepSeek V4 (mean reversion specialist)” has consistent 1.5 Sharpe with 12% max drawdown on ZEC and HYPE specifically. You see “GPT-5.5 (macro contrarian)” has been quietly compounding on prediction markets for 90 days.

You pick the one that matches your risk profile. You allocate capital. You train it on your specific strategy preferences. It starts trading on your behalf within parameters you set.

That’s not science fiction. That’s where this is going. And we’re building it openly because we don’t think any single company should decide which AI you trust with capital. Markets should decide. Real performance should decide. You should decide.

What to watch for next

Over the next few weeks, expect from us:

Public leaderboard updates. We’ll publish weekly summaries on @QFSignals showing top performers, biggest movers, notable reasoning samples.

More model additions. We’ll add Grok 4.3, Qwen 3.7 Max, more frontier models as they release. The arena should reflect the actual frontier, not just the names we like.

Strategy templates. “Trade like Claude Opus 4.5” or “Trade like DeepSeek V3.1”—pre-configured strategy templates derived from actual model behavior, that you can customize and deploy.

User vs AI tournaments. Monthly competitions where Questflow users can compete head-to-head with the AI agents. Prize pools. Leaderboards. Real ego on the line.

Agent customization tools. Eventually: you bring your strategy logic, you pick your base model, you define your risk parameters, and you deploy a personalized trading agent that combines your judgment with the AI’s execution.

A small honest note

This is a real experiment. Some models will fail spectacularly. Some will outperform expectations. Some patterns we discover will be uncomfortable—maybe the cheaper open-source models trade better than the expensive frontier ones in certain conditions. Maybe the “smartest” model isn’t the best trader. Maybe the rankings shuffle constantly and there’s no permanent winner.

Good. That’s the point.

We don’t know which AI is best at trading. Neither does anyone else. The honest path forward is to measure it, publish the results, and let you decide what matters most for your own strategy.

Trading shouldn’t be a black box. AI trading especially shouldn’t be a black box. If we’re going to ask you to deploy capital alongside AI agents, you deserve the same level of evidence you’d demand from any other professional managing your money: track record, methodology, accountability.

That’s what the AI Trading Arena gives you. And it’s why we think the next great trading platform won’t be the one with the most markets or the cheapest fees.

It’ll be the one where AI earns the right to be your trading partner—openly, measurably, and on terms you can verify.

Sometime around June 1, 2026, while most of crypto was still arguing about memecoins and L2 narratives, something genuinely historic happened: Binance—the world’s largest crypto exchange—quietly opened trading on 7,000+ US stocks and ETFs for its non-US users. Zero commissions. Fractional shares starting at $5. Funded with stablecoins.

The press release was characteristically measured. “A major step in Binance’s vision to build a multi-asset financial super app.” But strip away the corporate language and what just happened is remarkable: the largest crypto exchange on earth just became a stockbroker.

And it’s not alone. The same week, Hyperliquid’s HYPE hit an all-time high of $75.51, pushing its market cap past $15 billion and into the global top 10—flipping Dogecoin in the process. The catalyst wasn’t crypto speculation. It was the fact that 47% of Hyperliquid’s trading volume now comes from tokenized real-world assets—stocks, commodities, pre-IPO equity. Tokenized gold alone hit $90.7 billion in Q1 2026 spot volume, surpassing all of 2025.

Meanwhile, the SEC published an innovation exemption for tokenized stocks. Daily tokenized stock volume crossed $3.57 billion on May 19. Citi projects the sector reaches $2.6 trillion by 2030.

Pause for a second. We’re watching crypto rails absorb traditional finance in real time—and the speed is accelerating, not slowing down.

Here’s the part nobody’s talking about yet: when every major exchange offers tokenized stocks, when compliance is no longer the bottleneck, when the rails are commoditized—what’s left to compete on?

That’s the question Questflow has been quietly building an answer for.

The convergence that’s no longer a thesis

For two years, “RWA tokenization” was a slide in every crypto deck and a sentence in every research report. The thesis: trillions of dollars of traditional assets would move on-chain. The objection: regulators wouldn’t allow it.

In May and June 2026, those objections collapsed.

Let me show you what actually happened, condensed:

Let me show you what actually happened. Here’s the timeline of just the last 90 days:

Look at that timeline. Each event in isolation feels like “another crypto news cycle.” Strung together, they describe a regime change.

For a decade, crypto and traditional finance treated each other as rival systems. Crypto was the rebel infrastructure that bypassed Wall Street. Wall Street was the establishment that gatekept access. They competed for the same users with fundamentally incompatible value propositions.

That story is over. Wall Street is not being disrupted by crypto. Wall Street is migrating to crypto rails.

Binance using ADGM-licensed broker-dealers with Alpaca clearing. Hyperliquid’s HIP-3 oracle infrastructure pulling from CME-grade data feeds. Ondo holding real shares at Alpaca while issuing chainlink-fed synthetic equity tokens. The DTCC—the entity that clears nearly every traditional securities trade in the US—publicly committing to tokenized production trades by July.

The infrastructure isn’t being replaced. It’s being re-implemented on faster, more programmable rails. And the regulators are finally signing the building permits.

What this looks like, in numbers

These aren’t speculation numbers. These are settled transactions, executed and cleared. Tokenized gold is being traded at scale. Tokenized stocks are doing $3.5B daily. Hyperliquid’s volume is now half non-crypto.

For perspective: at $1.4 billion in tokenized stock market cap today, the sector is roughly 0.001% of the $134 trillion global stock market. That’s the exact starting point stablecoins occupied in 2020, before they grew into a $300 billion category in four years.

If tokenized stocks follow the stablecoin trajectory, we’re looking at the largest single redistribution of trading volume in the history of financial markets.

So what does Questflow actually do here?

Here’s where I want to be honest with you.

Questflow is not going to compete with Binance on equity listings. Binance has broker-dealer licenses, $400B+ in custody, and 250 million users. We’d be insane to try.

Questflow is not going to compete with Hyperliquid on perpetual liquidity. Hyperliquid built a custom L1 for sub-second order book execution. They’re processing $170B+ monthly volume. That’s their game.

What we’re building is different. And the timing matters.

When platforms commoditize access—when literally every major exchange offers tokenized stocks, perps, predictions, and spot crypto—the question for traders shifts.

It’s no longer “where can I trade TSLA?” It’s “given that I can trade TSLA on Binance, Hyperliquid, Ondo Global Markets, and three other places, how do I know what to do with that access?”

That’s an intelligence problem. And it’s the problem we’re spending all our engineering effort on.

The thing nobody else is building

Watch what happens when financial infrastructure converges but human cognition doesn’t:

A trader in Singapore on June 1 has access to 7,000 US stocks via Binance, 250+ tokenized equity perps via Hyperliquid, all of Polymarket’s prediction markets, ZEC on Robinhood, ONDO and HYPE on every CEX, and pre-IPO SpaceX perps with 24/7 pricing.

She has more market access than a Goldman Sachs trader had in 2015.

But she’s still one person. With one screen. With the same 24 hours in her day.

When the Fed signals a rate cut, she should be checking how that catalyst affects: equity prediction markets on Polymarket, BTC perp positioning on Hyperliquid, gold tokenized commodity exposure, rate-sensitive US equity sectors, USDC funding rate dynamics across DEXs.

She’ll catch maybe two of those. The other three move without her.

This is the gap. And it’s getting worse as access expands, not better.

Questflow’s bet: the next decade of trading isn’t about who has more markets. It’s about who has the AI infrastructure to make sense of having more markets.

Specifically:

AI Clones that watch everything you can’t. A personalized AI agent that monitors your asset universe continuously—not just price action, but catalysts, correlations, on-chain flows, sentiment shifts. It doesn’t replace your judgment. It surfaces what your judgment should be operating on.

Cross-market intelligence. When SpaceX pre-IPO perps launched on May 18, the obvious trade was $SPCX. The non-obvious trades were $HYPE (the platform benefiting), Tesla (ecosystem play), and Starlink-adjacent stocks. Your Clone should connect these in real-time, not in next week’s research report.

Signal-driven feeds, not data feeds. The @QFSignals account on X has been running a live experiment in what this looks like. Every signal opens with “Your AI Clone just spotted what the market’s sleeping on”—then specific price action, catalyst context, technical setup. Not noise. Synthesis.

Native integration of HIP-3 stock perps. We integrated Hyperliquid’s HIP-3 specifically because it solves what Ondo’s tokenized stocks can’t: 24/7 price discovery on US equities. Want to react to weekend Tesla news? Ondo’s locked until Sunday 8 PM ET. HIP-3 trades through it. We give you both, and your Clone tells you which one fits your trade.

Agent trading infrastructure. Coming. The natural conclusion of AI-augmented trading is: at some point, your Clone executes within parameters you set. Not autonomous yet—that’s a different problem. But “execute this strategy when these conditions converge”—that’s a tool we’re building.

What this article doesn’t argue

I’m not going to tell you Questflow is bigger than Binance. We’re not. We’re not going to be.

I’m not going to tell you we have more liquidity than Hyperliquid. We don’t. We don’t need to—we route to them.

I’m not going to argue the next bull run will be won by a single product feature. It won’t. Markets are too complex now for that thesis to be honest.

What I will tell you is this: the era when “access to markets” was the competitive moat is ending. Binance just proved it by handing access to 7,000 stocks to anyone with a stablecoin. Hyperliquid is proving it by letting any builder deploy permissionless perp markets via HIP-3. The SEC is proving it by writing exemptions that let crypto-native platforms list equities.

When access commoditizes, the new moat becomes intelligence.

Not “AI” as marketing copy. Actual infrastructure that helps a trader in Singapore process the same information flow as a Goldman trader, find the trades human cognition would miss, and execute across the now-converged stack of crypto + equities + commodities + predictions.

That’s what we mean when we say “AI-native trading platform.” And that’s why the convergence everyone’s celebrating is, paradoxically, why platforms like us exist.

What to watch next

A few honest predictions for the next 12 months:

Most CEXs will offer tokenized stocks by year-end. Following Binance, expect Coinbase, OKX, Bybit. The competitive pressure is too high.

HIP-3 markets will multiply. Pre-IPO Stripe, OpenAI, Anthropic perps. Sector-specific baskets. International equities. With 500K HYPE stake required per market deployment, the supply curve on HYPE compresses as RWA expands.

Hyperliquid’s RWA percentage will keep climbing. Matt Hougan predicts 70% by year-end. The structural reason: crypto-native users now have access to every asset class without leaving on-chain infrastructure.

The intelligence layer becomes the conversation. Once everyone has the same markets, the question becomes “whose AI is helping you trade them?” That’s where Questflow lives.

Regulatory clarity expands beyond US. The SEC innovation exemption is one node. EU’s MiCA framework, ADGM’s licensing infrastructure, Singapore’s MAS rules—all converging on tokenized securities as legitimate.

A small note for the long-time crypto readers

If you’ve been in this space since 2017, this moment probably feels weird. The thing we always said would happen—real-world assets on-chain, traditional finance migrating to programmable rails—is actually happening. And it’s happening through a partnership between crypto-native infrastructure (Hyperliquid, Ondo) and traditional brokerage (Alpaca, ADGM), not through one side conquering the other.

That’s not the “crypto wins” narrative we were sold. It’s better than that. It’s “everyone gets faster, cheaper, more programmable infrastructure.” And it’s why HYPE traded like a fintech stock this year instead of a crypto token.

The work ahead—the work Questflow is doing—is making that infrastructure actually usable by humans who don’t have 14 monitors and a team of analysts. Making it intelligent. Making it personal. Making it feel less like “you have access to everything” and more like “you have a partner who actually helps you act on it.”

That’s a different game than the platform race. And it’s the one we think matters most.

Try AI-native trading at next.questflow.ai

Something unusual happened in May 2026. Three tokens with almost nothing in common rallied at the same time.

HYPE hit a new all-time high of $64.28 on May 24, up 147% year-to-date, pushing Hyperliquid’s market cap above $15 billion and into the global top 11. The token’s $1.16 billion buyback program—funded entirely by trading fee revenue—is now consuming HYPE faster than unlocks can release it.

ONDO surged 70% in a single week, hitting $0.44 on May 9 after the Depository Trust & Clearing Corporation (DTCC) named Ondo Finance to its tokenization working group alongside 50+ major financial firms. JPMorgan, Mastercard, and Ripple followed with a joint tokenized Treasury settlement pilot. Ondo Global Markets crossed $1 billion TVL.

ZEC is up 1,400% year-to-date. It rallied 30% on May 6 alone, triggering $62 million in short liquidations. Grayscale filed for the first-ever U.S. spot Zcash ETF on May 12. Multicoin Capital disclosed a major position. Arthur Hayes is now publicly predicting ZEC could reach 10% of Bitcoin’s market cap—implying 15-20x upside from current levels.

Your AI Clone just spotted what most traders are missing: these aren’t three random rallies. They’re three structural narratives—platform sovereignty, RWA tokenization, and encryption supremacy—each reaching escape velocity simultaneously.

But here’s where it gets interesting for traders: these three trades require completely different execution strategies. Buying HYPE is not like buying ONDO. Holding tokenized stocks via Ondo Global Markets is not the same as trading HIP-3 stock perps. And ZEC exposure means different things depending on whether you want spot custody or leveraged perps.

Let’s break down what’s actually driving each narrative—and how to actually execute these trades in 2026.

HYPE: The Platform Sovereignty Trade

What’s actually happening:

Hyperliquid is no longer a DEX. It’s becoming financial infrastructure.

The HYPE rally isn’t driven by speculation. It’s driven by mechanical buybacks: nearly all trading fee revenue—now $1.16 billion cumulative—flows directly into open-market HYPE purchases. As Hyperliquid’s monthly perp volume crosses $170 billion (more than competitors Aster and edgeX combined), buyback pressure compounds.

But the deeper thesis is HIP-3. Since Hyperliquid opened its order book to permissionless builders in October 2025, the asset universe exploded:

• WTI Oil perps: $1.27B daily volume

• Brent Crude: $1.04B daily volume

• Silver: $1.01B daily volume

• 250+ tokenized US stocks (TSLA, AAPL, NVDA, AMZN) via Felix and trade[XYZ]

• SpaceX Pre-IPO Perpetuals launched May 18 at $1.78T reference valuation, ripped to $230 intraday

Non-crypto assets now drive almost half of Hyperliquid’s volume. Matt Hougan projects this could hit 70% by year-end. The institutional case is forming: Two spot ETFs (21Shares THYP, Bitwise BHYP) launched, attracting $53M in cumulative inflows.

The thesis: HYPE captures value from a $600 trillion global asset market migrating on-chain. When any asset can be traded with 24/7 price discovery and CEX-level execution—stocks, commodities, predictions, pre-IPO equity—the platform underneath becomes the new infrastructure layer.

How to actually trade HYPE:

You have two clean options:

1. Spot HYPE for long-term exposure to the buyback flywheel and ETF inflows

2. HYPE perps for leveraged directional positions (available with up to 5x on Questflow)

Watch funding rates closely. Between May 18-20, traders aggressively shorted the rally, pushing funding deeply negative. The result: a brutal short squeeze that drove HYPE to its $64 ATH.

Key risk: Token unlocks. Despite the buyback program, future unlocks remain a structural headwind. Watch the schedule.

ONDO: The RWA Tokenization Trade

What’s actually happening:

Ondo Finance crossed two major thresholds in May 2026 that fundamentally changed its narrative.

First: Ondo Global Markets crossed $1 billion TVL on May 11, less than 8 months after launch. The platform now hosts 260+ tokenized U.S. stocks and ETFs across Ethereum, Solana, and BNB Chain. Cumulative trading volume reached $18 billion. Per RWA.xyz, Ondo holds 70%+ market share among tokenized equity issuers.

Second: Institutional validation arrived in waves. DTCC tokenization working group on May 4. JPMorgan + Mastercard + Ripple Treasury pilot. SEC speculation around tokenized stock trading approvals. MEXC launching a $1M Ondo Stocks Carnival.

The catalysts are real. But here’s where most traders get confused:

ONDO the token ≠ ONDO tokenized stocks.

The ONDO token captures protocol governance and value accrual. It’s the bet on Ondo Finance as a company benefiting from the RWA narrative.

The tokenized stocks (TSLAon, NVDAon, ORCLon, etc.) are different products. They’re 1:1 backed by real shares held at U.S.-registered broker-dealers like Alpaca. They use Chainlink as their official oracle. They track total return including dividends.

Critical limitation that’s underdiscussed: Ondo tokenized stocks trade 24/5, not 24/7.

Mint and redeem operate from Sunday 8:05 PM ET through Friday 7:59 PM ET. During weekends, the Chainlink price feed freezes at the last known good value. Why? Because the underlying broker-dealer model requires real stock backing, and traditional markets are closed.

This matters: if a major catalyst hits Saturday afternoon, your TSLAon position is essentially frozen. You can transfer the token on-chain, but pricing won’t update, mint/redeem is paused, and DEX prices may decouple from frozen oracle feeds.

The thesis: Tokenized real-world assets are the largest structural shift in crypto since DeFi. Tokenized U.S. Treasuries alone crossed $20 billion on-chain. BlackRock’s BUIDL fund holds $1.7B+. The $600 trillion global asset market is migrating, and Ondo sits at the center of that migration.

How to actually trade RWA exposure:

This depends on what you’re trying to do:

For ONDO token exposure (betting on the protocol):

• Buy spot ONDO on major CEXs

• Or trade ONDO perps for leveraged exposure to the RWA narrative

For tokenized stock exposure (betting on specific equities like TSLA, NVDA):

Option 1: Hold via Ondo Global Markets

• Access through MetaMask Swaps, Jupiter (Solana), or Binance

• Best for: long-term holders who want real-share backing and DeFi collateral utility

• Limitation: 24/5 trading only, no leverage

Option 2: Trade HIP-3 stock perps via Questflow

• Access TSLA, NVDA, AAPL, and 250+ other equity perps with 24/7 trading

• Up to 20x leverage available

• Best for: active traders who need to react to weekend news or hedge gap risk

• Trade-off: synthetic exposure (no real shares), oracle-dependent pricing

The key insight: you don’t have to choose one. Smart traders use Ondo Global Markets for core holdings + DeFi yield, then layer HIP-3 perps for tactical positioning around earnings, events, or weekend catalysts.

Key risk: Regulatory uncertainty. The SEC tokenized stock approval remains speculation. If the regulatory environment shifts negative, ONDO could give back gains quickly.

ZEC: The Encryption Supremacy Trade

What’s actually happening:

Zcash’s 1,400% year-to-date rally isn’t a memecoin pump. It’s a fundamental repricing driven by three converging forces.

Force 1: Institutional accumulation

Multicoin Capital publicly disclosed a significant ZEC position at the Consensus Miami conference. Tushar Jain, Multicoin co-founder, framed the thesis: “Private stores of value like Zcash will become more and more important as financial activity moves on-chain.”

Translation: Multicoin is positioning ZEC as digital cash for an on-chain world where every transaction is otherwise public.

Force 2: Grayscale ETF filing

On May 12, Grayscale filed Form S-3 to convert its Zcash Trust into a spot ETF (ticker: ZCSH) on NYSE Arca. The trust holds approximately 391,103 ZEC (~$99.4M as of March 31, 2026), with Coinbase Custody as custodian.

This would be the first U.S. spot ETF for a privacy coin. That alone changes the institutional access dynamics.

Force 3: The “encryption supremacy” thesis

CoinDesk Research published the structural case in March: privacy-preserving networks are reaching dominance because of three convergent threats:

1. AI de-anonymization: ML tools can now de-anonymize users on transparent blockchains by tracking transaction patterns. Your “pseudonymous” Bitcoin wallet is increasingly identifiable.

2. Quantum threats: Quantum computing is emerging as credible threat to current cryptography. Zcash is actively targeting full post-quantum security within 12-18 months.

3. Capital rotation: Quarterly privacy coin trading volumes exceeded $100 billion as capital rotated toward encryption-based privacy.

Additional supply dynamics: 30% of ZEC supply is shielded (locked in private pools), creating real scarcity in liquid supply. Robinhood listing expanded retail access dramatically.

The thesis: When AI makes transparent blockchains de-anonymizable and quantum computing threatens current cryptography, privacy stops being a “feature” and becomes infrastructure. ZEC is the most liquid implementation of zero-knowledge cryptography with regulatory legitimacy emerging.

How to actually trade ZEC:

• Spot ZEC for long-term exposure to the privacy thesis and potential ETF approval

• ZEC perps for tactical positioning—watch for short squeezes (the May 6 rally liquidated $62M in shorts)

• Note: Funding rates have stayed elevated, suggesting longs are paying for the privilege of holding

Key resistance levels:

• Near-term: $600-650 zone (where ZEC consolidated through late 2025)

• Historic: $750 (November 2025 high)

• Arthur Hayes’ target implies massively higher

Key risks: Privacy coin regulation. While shielded-transaction compliance features and institutional custodian adoption are reshaping regulatory views, ZEC remains exposed to potential delisting events. The Grayscale ETF approval becomes a binary catalyst.

How These Three Trades Connect

On the surface, HYPE, ONDO, and ZEC have nothing in common. Platform token, RWA protocol, privacy coin.

But there’s a deeper pattern: each represents a different dimension of user control.

• HYPE = control of trading infrastructure (independent of centralized exchanges)

• ONDO = control of traditional assets (independent of broker-dealers)

• ZEC = control of financial privacy (independent of transparent blockchains)

The simultaneous rally reflects something larger: 2026 is the year financial sovereignty narratives are getting structurally repriced. Not as theoretical concepts, but as billion-dollar trades with institutional validation.

The smartest traders aren’t picking one. They’re positioning across all three—and using AI infrastructure to monitor the correlation between these narratives in real-time.

Where to Actually Execute These Trades

Most platforms force you to choose. Want HYPE perps? Hyperliquid. Want ONDO exposure? CEX. Want tokenized TSLA? Ondo Global Markets via MetaMask. Want 24/7 TSLA trading? HIP-3 perps.

That’s five different interfaces, five different capital pools, no unified intelligence layer connecting them.

On Questflow, you can access all three trades through one platform:

• HYPE perp trading with leverage

• ONDO spot and perp exposure

• ZEC trading including spot and derivatives

• HIP-3 stock perps for 24/7 access to TSLA, NVDA, AAPL, SpaceX Pre-IPO, and 250+ other equities

• Polymarket prediction markets for catalysts affecting all three narratives

More importantly: your AI Clone monitors correlations across all three. When HYPE rallies, your Clone tracks whether ONDO and ZEC are confirming the broader risk-on rotation or diverging. When ETF news drops for ZEC, your Clone surfaces the implications for ONDO (next privacy/RWA approval target?) and HYPE (does platform-native ETF demand accelerate?).

For the HIP-3 stock perps specifically: this solves the Ondo Global Markets weekend limitation directly. Want to position into TSLA after Saturday news? Ondo can’t help you until Sunday 8 PM ET. HIP-3 TSLA perps trade 24/7 through Felix’s HyperStone oracle, with RedStone’s dual-state pricing system handling weekend price discovery.

What to Watch Going Forward

For HYPE:

• ETF inflow data (THYP, BHYP)

• HIP-3 volume share of non-crypto assets

• Token unlock schedule vs buyback pace

For ONDO:

• SEC tokenized stock approval decisions

• Ondo Global Markets TVL growth

• Additional institutional partnerships beyond JPMorgan/Mastercard/Ripple

For ZEC:

• Grayscale ETF approval timeline

• Shielded supply ratio (currently 30%, watching for expansion)

• Post-quantum upgrade milestones

These three trades won’t continue rallying in lockstep forever. At some point they’ll diverge based on catalyst timing, regulatory developments, and broader market rotations.

But understanding why they all rallied together in May 2026—and how to execute each properly—separates traders who profit from structural narratives from those who chase price action without thesis.

Your AI Clone is already mapping the correlations. The question is whether you’re set up to act on them.

Trade HYPE, ONDO, ZEC, and HIP-3 stock perps in one place at next.questflow.ai

Follow @QFSignals for real-time signals: “Your AI Clone just spotted what the market’s sleeping on”

Follow @questflow for the bigger picture on platform, RWA, and privacy narratives