Artificial Finance

AI in Finance

A 27B Model Just Beat GPT-5.6 at Finance Research, for 78% Less

A new arXiv preprint reports a 27-billion-parameter finance-native agent outscoring GPT-5.6-Sol and Claude Opus 4.8 on seven financial benchmarks at roughly a fifth of the cost. What it means for research budgets, model risk and regulation.

By The Artificial Finance Desk 12 min read

A technical report appeared on arXiv with a claim that, if it survives outside its authors' own test harness, changes the arithmetic of investment research. A 27-billion-parameter model, small enough to run on one server, reportedly outscored GPT-5.6-Sol and Claude Opus 4.8 on all seven professional financial benchmarks its authors tested, at roughly a fifth of the cost per task.

The interesting part is not the leaderboard. The authors did not train the model to give better financial answers. They trained it so that every answer carries a recoverable trail from source document to calculation to conclusion, then built a reward function that pays nothing when that trail breaks.

What happened#

The paper, Mint-Agent: Introducing Finance-Native Agentic Foundation Models, was submitted on 17 August 2026 and has not been peer reviewed. It describes two models: Mint-Cu at 9 billion parameters, built on Qwen3.5-9B, and Mint-Ag at 27 billion, built on Qwen3.6-27B.

The reported results are broad. Mint-Ag scores 91.33 per cent on FinanceBench, a disclosure-grounded question-answering set, and 55.71 per cent on BizFinBench. On agentic tasks, where the model must go and find the evidence itself, it reports 76.00 per cent on FinanceAgentBench v1.1, 60.49 per cent on v2, and 89.04 per cent on Task 2 of FinSearchComp. Claude Opus 4.8 scores 66.00 per cent on FinanceAgentBench v1.1 in the same table; GPT-5.6-Sol scores 41.10 per cent on FinSearchComp T2.

Cost is where this gets awkward for incumbents. On FinanceAgentBench v2 the paper puts Mint-Ag at $0.213 per task against $0.959 for GPT-5.6-Sol, a 77.8 per cent reduction alongside a 3.70 point accuracy gain. The 9B model answers v1.1 tasks at $0.016 each. Read the footnote, though: these are not metered API bills. The authors estimated cost by applying published base-model prices to measured token usage on locally hosted checkpoints.

Why auditability is an engineering problem, not a compliance slogan#

A large language model predicts text. An agent is a model given tools and a loop: it can search, open a filing, run a calculation, look at the result and decide what to do next. A long-horizon task runs that loop for dozens of steps. One case study in the paper is a 37-step investigation of TSMC revenue that hits blocked filings and wrong-quarter pages before recovering.

Financial research breaks agents in a particular way. The answer is rarely one fact. It is a fact read under the correct fiscal period, reconciled against a second source, converted into the right units, then carried into a calculation. Any link can fail quietly, and a figure pulled from the wrong quarter looks exactly like one pulled from the right quarter.

The response here is architectural. The harness keeps an evidence ledger outside the model's context window: a persistent, typed record of every fact retrieved, with its source locator, entity, reporting period, metric, value and unit. Failed retrieval paths are logged separately, which keeps "no support exists" distinct from "the source was unreachable". A compact working memory carries the plan and open questions across turns.

Training data comes from primary sources rather than existing question banks: SEC EDGAR filings and issuer archives, values aligned to their XBRL tags, exchange records, and series from FRED. XBRL is the tagging standard that lets "revenue" in a filing match a defined accounting concept instead of a string of text. A synthesised task is discarded unless its derivation replays against that evidence and reaches a unique answer.

Where this lands in markets#

Incumbents are already selling into this workflow. LSEG has put a Deep Research agent into Workspace, drawing on its own pricing and time-series data. JPMorgan told CNBC in June it would deploy agents that work autonomously for far longer than earlier versions, and Morgan Stanley is opening its wealth management funnel to agents. FCA chief executive Nikhil Rathi told an industry audience on 24 June that more than 80 per cent of financial services firms already use AI, and called agentic systems in wholesale markets "a profound step change to the structure and operation of markets".

If a 9B model does useful disclosure research at under two cents a task, the binding constraint stops being cost per query and becomes the cost of verifying output. Spending moves from inference towards controls, data licensing and review capacity. It touches the capital expenditure debate too. The BIS Annual Economic Report, published on 28 June, notes that the five largest hyperscalers plan to spend more than $1 trillion on AI capacity through 2026, outpacing earnings and free cash flow, while aggregate productivity gains look modest at under 1 per cent over a long horizon. Evidence that specialised small models capture much of the value in one vertical is a data point in that debate, not a resolution of it. Any read-through to equity valuations is interpretation, not forecast.

For quantitative teams the implication is a new dependency. If research inputs come from agents, provenance becomes part of the data pipeline, and the ledger that satisfies an auditor is what lets a backtest reconstruct what was knowable at a point in time. Get that wrong and you have industrialised look-ahead bias.

Inside the twin-specialist training stack#

The recipe matters because it is becoming a standard pattern. Two specialists are trained from a shared base model. One handles bounded reasoning: extraction, calculation, verification against supplied evidence. The other handles execution: search, tool use, knowing when to stop.

Both begin with supervised fine-tuning on filtered traces from larger teacher models, then move to reinforcement learning with verifiable rewards, where a program computes the reward rather than a human or a judge model. The verifier checks that the answer matches, that every premise traces to a source locator, and that the derivation replays under its unit and period constraints. An answer that looks right but cannot be reconstructed earns nothing. Optimisation uses Group Sequence Policy Optimization, which scores whole sequences rather than single tokens.

The execution specialist gets an extra stage. Failed trajectories are inspected for critical steps, single turns where a different action would have restored a valid route to the answer, and only those decisions are corrected against a frozen teacher model. That is far cheaper than imitating entire expert trajectories.

Merging the two is where such projects usually lose their gains. The authors apply TIES-Merging, a peer-reviewed NeurIPS 2023 method that trims small parameter changes and resolves sign conflicts between fine-tuned models. Merging alone cost 3.86 points on BizFinBench. A second stage, multi-teacher on-policy distillation, routes each example to whichever specialist should teach it while the student generates the states, recovering 1.33, 3.86, 4.00 and 3.19 points across four benchmarks.

What the scores do not prove#

This is a self-reported preprint, and the authors ran both Mint-Agent and the frontier baselines on FinSearchComp inside their own harness. A harness built around one model's assumptions is not a neutral instrument.

The test sets are small. FinanceAgentBench was evaluated on 50 public v1.1 tasks and 27 public v2 tasks, because no held-out split is publicly available and, by the authors' account, the maintainers had not replied to a request for one. Twenty-seven tasks is a thin base for a 3.70 point claim. For scale, the Finance Agent Benchmark preprint built 537 expert-authored questions across nine task categories and reported OpenAI o3 as the best model at 46.8 per cent accuracy, at $3.79 per query.

The paper's own failure analysis is the most useful section in it. On FinanceAgentBench v2 the 9B model reaches 41.98 per cent, and its dominant error is evidence extraction at 18.5 per cent: the source is reached, but the relevant fact is not captured with the correct scope. That is exactly the failure a portfolio manager cannot spot by reading the output.

Then there is the regulatory gap. The Federal Reserve, OCC and FDIC issued SR 26-2 on 17 April 2026, the first revision to US model risk guidance in fifteen years, superseding SR 11-7, and it explicitly places generative and agentic AI outside its scope. Singapore went further: MAS and industry partners published Safeguards for Agentic Finance at Runtime on 3 July 2026, setting out how agent actions are authorised and what is recorded at each decision.

There is one more caution. The Financial Stability Board has warned that correlated AI approaches can amplify volatility and raise the probability of flash crashes, with a market segment running the same models and data behaving as if it were a single institution. Cheap specialised models make that convergence more likely, because cheapness invites everyone to adopt the same one.

From SR 11-7 to agents that cite their sources#

Finance has done model governance before. The 2011 regime treated a model as a static artefact: documented, validated, monitored, owned. That worked because a credit scorecard does not go looking for its own inputs.

An agent does. It selects sources, judges when evidence is sufficient, and terminates on its own assessment. Validating that is closer to auditing a junior analyst than validating a regression, which is why SR 26-2 carved agentic systems out instead of stretching the old frame over them.

Read that way, Mint-Agent is less a modelling breakthrough than a structural change in what "the model" means. The claim under test is that auditability can be trained rather than bolted on. If it replicates, the moat in financial AI shifts from parameter count towards licensed data, provenance infrastructure and verification. If it does not, these numbers will be remembered as a harness artefact.

Key takeaways#

  1. A preprint dated 17 August 2026 reports a 27B finance-tuned agent beating GPT-5.6-Sol and Claude Opus 4.8 across seven financial benchmarks, at 72.5 per cent below the comparison mean on FinanceAgentBench v1.1 and 77.8 per cent below GPT-5.6-Sol on v2. It is self-reported and not peer reviewed.
  2. An answer counts as correct only if its evidence trail replays, which makes auditability a training decision rather than a reporting feature.
  3. Cost figures are estimates from base-model pricing applied to token counts, not metered API spend.
  4. The public splits used were 50 and 27 tasks. A few points of advantage on 27 items is fragile.
  5. US supervisors left agentic AI outside revised model risk guidance under SR 26-2, while MAS published a runtime safeguards framework. Firms deploying these systems are ahead of their rulebook.

Frequently asked questions#

What is a finance-native agentic model? A model post-trained for financial work, with tools for searching filings and market data plus a persistent record of what it retrieved and from where. The alternative is a general-purpose model handed financial tools at inference time.

Does a smaller model really beat a frontier model? On these benchmarks, as reported by the authors, yes. Specialisation buys a lot inside a narrow domain. It says nothing about general capability, and part of the evaluation ran on the authors' own harness.

What is RLVR? Reinforcement learning with verifiable rewards. A program scores the output instead of a human or judge model: does the answer match, does every claim trace to a source, does the calculation replay under the right units and period.

Why does 27 tasks matter so much? A 3.70 percentage point difference across 27 items is roughly one task. Small evaluation sets produce large apparent gaps that rarely survive a bigger sample.

Can these systems be used in a regulated workflow today? SR 26-2 puts generative and agentic AI outside its scope and points firms to their own risk practices. That is permission to proceed carefully, not a validated pathway. MAS SAFR is the most detailed public template for runtime controls.

What should a buy-side team measure before adopting one? Focus less on final answer accuracy and more on how often retrieved evidence is captured with the correct scope, period and units. That is where the 9B model failed most often on the harder benchmark.

References#

Primary research

Regulators and central banks

Industry and market sources

This article analyses published research and policy documents. It is not investment advice and makes no recommendation about any security, vendor or trading strategy. Benchmark results described here are self-reported by the authors of the cited preprint and have not been independently replicated.

Related coverage

AI in Finance

India's Central Bank Just Redrew the Rules of the AI Banking Race

RBI Governor Sanjay Malhotra told bankers in Mumbai that the winners of the AI era will be those who understand what they deploy, not those who move fastest. Here is why his warning about model concentration matters for global markets.

12 min read