The best AI analyst scores 65.4% on a Wall Street research test

If you want an AI to do your earnings homework, the first question is simple: how often does it get the answer right? The Finance Agent v2 benchmark from Vals AI is one of the few public leaderboards that tries to measure exactly that, on questions built from real company filings.

The scoreboard

As of October 12, 2026, the public Finance Agent v2 leaderboard lists 78 AI models. Google's Gemini 4 Argon sits at the top with 65.40% accuracy, followed by Gemini 3.8 Flash at 61.44% and Meta's Muse Spark 1.2 at 60.60%. Muse Spark 1.3 Max is fourth at 59.96%. nContext: n=450 test questions, 3 runs · Oct 12, 2026 leaderboard · Vals AI · ±0.32 SEM for the leader, ±0.13 and ±0.28 for the next two

Below the leader, the pack is tight: second place and tenth place are separated by less than three points. The best OpenAI entry on the board, GPT-5.6 Luna, ranks 22nd at 55.04%. The best xAI entry, Grok 4.6, ranks 30th at 53.68%. At the bottom of the table, Command A+ scores 9.04%. nContext: n=450 test questions, 3 runs · Oct 12, 2026 leaderboard · Vals AI · ±0.31 (GPT-5.6 Luna), ±0.67 (Grok 4.6); other margins not shown in the table

What the test actually asks

Finance Agent v2 contains 927 expert-reviewed questions split into three parts: 27 public samples, 450 private validation questions that Vals licenses, and 450 held-out test questions. The published scores come from the held-out test set only, which makes it harder for model makers to train on the answers.

Vals says the questions target the depth expected of a second- or third-year investment banking analyst. Examples on the benchmark page include comparing capital allocation at Walmart, Costco and Target, or ranking major U.S. banks by excess CET1 capital over their regulatory minimums.

Every model gets the same six tools: an SEC EDGAR search, web search, an HTML page parser, a retrieval tool over fetched pages, a calculator and a price history tool. Each task has a two-hour time limit, and every model is run three times, with the mean reported.

Grading is strict. Each question has weighted checks, and some are flagged as dealbreakers: miss one of those and the whole answer scores zero. A jury of three AI models (GPT-5.4, Gemini 3.1 Pro and Claude Sonnet 4.6) does the grading.

Where the AI breaks

The category breakdown is the most useful part for traders. Reading and summarizing is largely handled; building numbers across several documents is not. nContext: n: not disclosed per category · Oct 12, 2026 leaderboard · Vals AI · margin: not disclosed

• Earnings analysis: best score 84.8%
• General qualitative analysis: 83.4%
• General quantitative analysis: 81.8%
• Market analysis: 79.7%
• Disclosure analysis: 74.5%
• Adjustments (GAAP to non-GAAP bridges): 60.5% in Vals' category chart, 56.29% in its written summary. The source disagrees with itself, so treat this one as unsettled.
• Comparables: 52.0%
• Financial modeling (DCF, LBO, accretion and dilution): 34.5%

In Vals' own words, models handle simple retrieval but "still struggle to perform reliably on harder, multi-step financial work that relies on precise numbers and specific industry convention."

Paying more does not buy much more

Cost per test varies far more than accuracy. Muse Spark 1.2 scores 60.60% at $0.77 per test, less than a fifth of the $4.38 the leader costs. GLM 5.3 Flash scores 57.85% at $0.05 per test. Speed differs too: Gemini 4 Argon averages 16 minutes 21 seconds per task, against 3 minutes 21 seconds for Gemini 3.8 Flash. nContext: n=450 test questions, 3 runs · Oct 12, 2026 leaderboard · Vals AI · ±0.28 (Muse Spark 1.2), ±2.06 (GLM 5.3 Flash)

Vals draws the same conclusion from its cost chart: mid-tier models sit a few points below the leaders at a fraction of the price, so the accuracy premium at the frontier is modest compared with the price gap.

How fast the bar is moving

The original Finance Agent Benchmark paper, published in May 2025, reported that the best model it tested, OpenAI's o3, reached 46.8% accuracy nContext: n=537 questions in the full v1 dataset · paper of May 20, 2025 · arXiv:2508.00828 · margin: not disclosed in the abstract at an average cost of $3.79 per query. That was version 1 of the test, with looser grading and fewer tools, so the numbers are not directly comparable. Still, the direction is clear: the newer version is harder, and the top score is higher.

How we will track it

The AI Analyst Scoreboard will come back to this leaderboard and to other published rankings of AI analysts. We relay the scores as published, link the source, and say what changed. We do not run our own tests.