Finance Agent v2 — Results
| Rank | Model | Score |
|---|---|---|
| 1 | gemini-4-argon | 65.4 |
| 2 | kimi-k3 | 54.4 |
Finance Agent v2
1 phaseActive
Vals AI benchmark of 927 expert-reviewed questions on public company filings across 9 analyst task types. The leaderboard uses a private held-out test split. Metric: % accuracy.
Overview
Finance Agent v2
Quick answer: Finance Agent v2 is Vals AI's benchmark for the work of an entry-level financial analyst: answering hard questions about public company filings with search and document tools. It has 927 expert-reviewed questions in 9 task types. Gemini 4 Argon leads at 65.4% (Vals AI, October 2026).
At a Glance
What it tests: Whether an agent can find, read and reason over public company filings to answer the questions an entry-level analyst gets, from simple lookups to modeling.
Why it matters: It is one of the few finance benchmarks built from real analyst work and graded by domain experts, and the leaderboard is run by an independent evaluator rather than reported by model developers.
Known limitations: The leaderboard is scored on a private held-out test split, so outside teams (Benchgen included) cannot reproduce the numbers; they can only be mirrored from Vals AI. Vals reports two accuracy views (see below), and the headline number depends on which one is quoted.
What Finance Agent v2 Measures
The questions cover 9 task types: earnings analysis, general qualitative, general quantitative, market analysis, disclosure analysis, adjustments, comparables, precedents and financial modeling. The agent works over public company filings and has to retrieve the right documents, extract the numbers and reason to an answer. Version 2 builds on Finance Agent v1.1 (537 questions) with more questions, 9 task types and expert review of every question.
Benchmark Specifications
| Field | Value |
|---|---|
| Task category | Agent / finance analysis |
| Questions | 927, split into public, private validation and held-out test sets |
| Task types | 9 (earnings, qualitative, quantitative, market, disclosure, adjustments, comparables, precedents, modeling) |
| Metric | % accuracy (partial credit; all-pass also reported) |
| Test set | Private held-out split |
| Saturation | Low |
| Created by | Vals AI |
| Predecessor | Finance Agent v1.1 (537 questions) |
| Leaderboard | vals.ai/benchmarks/fabv2 |
How Finance Agent v2 Is Scored
Vals AI reports accuracy two ways: partial credit, the leaderboard default, and all-pass, where an answer counts only if it is fully correct. The top partial-credit score is 65.4%; the top all-pass score is 50.9%. Scores on this page are partial credit unless marked otherwise.
State-of-the-Art Results
Vals AI leaderboard, partial credit, updated 2026-10-12:
| Rank | Model | Score | Source | Date |
|---|---|---|---|---|
| 1 | Gemini 4 Argon | 65.4% | Vals AI | 2026-10 |
| 2 | Gemini 3.8 Flash | 61.4% | Vals AI | 2026-10 |
| 3 | Muse Spark 1.2 | 60.6% | Vals AI | 2026-10 |
| 4 | Muse Spark 1.3 Max | 60.0% | Vals AI | 2026-10 |
| 5 | Wonder Full | 59.1% | Vals AI | 2026-10 |
Also on the Benchgen leaderboard: Kimi K3 54.4%, as reported by Moonshot AI in July 2026.
Finance Agent v2 on Benchgen
No Benchgen runs. The test set is private, so the scores above are mirrored from Vals AI and model developers and cannot be re-run on Benchgen.
Finance Agent v2 vs Other Benchmarks
| Benchmark | What it tests | Saturation |
|---|---|---|
| Finance Agent v2 | Analyst questions over public company filings | Low |
| CorpFin v2 | Long credit agreements (archived by Vals AI) | High |
| SpreadsheetBench 2 | Spreadsheet manipulation & formula tasks | Low |
| SaaS-Bench | SaaS product usage workflows | Low |
Run Finance Agent v2 on Your Model
The leaderboard test set is private, so you cannot reproduce the official score yourself. Vals AI publishes a public split, and the v1 harness is open on GitHub; on Benchgen you can build a comparable evaluation from your own filings and questions and track your agent over time.