Vals AI, 2025

Finance Agent Benchmark

Checks whether a tool-using agent can do entry-level financial analyst research, such as answering hard questions from public company filings. Question types include quantitative and qualitative analysis, comparables, adjustments, earnings, disclosures and modelling.

Best reported

67.8

Claude Opus 5.5Anthropic

Partial Credit accuracy % (v2.1)

Official board, 29 Oct 2026

About this benchmark

Size
v2.1: 926 questions (27 public, 450 private validation, 449 test). v1 paper: 537 questions.
Public split
27 public questions plus the agent harness in the GitHub repo. The v1 paper released 50 public validation questions (CC BY 4.0).
Grading
Each question has weighted rubric checks, some marked as dealbreakers. A missed dealbreaker scores the question zero. Headline metric is Partial Credit; All-Pass is the secondary. A three-model LLM jury grades (GPT-6.1 Sol, Gemini 3.8 Flash, Claude Sonnet 5.5). Two-hour limit per task.
Licence
MIT (code, finance-agent-v2 repo). The v1 paper lists CC BY 4.0 for the public data and MIT for the code.
Submission
Not stated
Products allowed
Not stated
Latest version
v2.1
What it does not cover
Does not test trading, portfolio decisions, advice suitability or spreadsheet model building.

Results

10 models
Leaderboard

Versions are not comparable. Vals publishes Finance Agent v1.1 and v2.1 as separate leaderboards on different question sets. Scores from different versions are not comparable, so each version is its own board here; this page shows v2.1.

Each score shows who reported it: the benchmark's authors (quoted from their paper), the benchmark's official board, or Moneda (run with the benchmark's own grader and rules). We do not copy other aggregators.

Best value: DeepSeek V4.1 Flash, 65.4 at $0.370 per question (cheapest within 5% of the top score)

Finance Agent Benchmark: score vs cost

Score against cost per test. Up and to the left is better.

Updated 29 Oct 2026

Y: Partial Credit accuracy % (v2.1)X: Cost per test (USD)Showing 10 models
Finance Agent Benchmark: score vs cost10 models. Frontier: DeepSeek V4.1 Flash, Gemini 3.8 Flash, Wonder Full, Gemini 4 Argon, Claude Opus 5.5.65666768$0$2$4$6$8EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Claude Opus 5.5: 67.8 Partial Credit accuracy % (v2.1), $8.91 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5.5Gemini 4 Argon: 67.7 Partial Credit accuracy % (v2.1), $4.32 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 ArgonWonder Full: 67.6 Partial Credit accuracy % (v2.1), $2.68 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Wonder FullClaude Sonnet 5.5: 67.3 Partial Credit accuracy % (v2.1), $6.95 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Sonnet 5.5Gemini 4 Carbon: 67.2 Partial Credit accuracy % (v2.1), $7.29 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 CarbonGemini 3.8 Flash: 65.8 Partial Credit accuracy % (v2.1), $1.99 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 3.8 FlashWonder: 65.5 Partial Credit accuracy % (v2.1), $2.62 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)WonderDeepSeek V4.1 Flash: 65.4 Partial Credit accuracy % (v2.1), $0.37 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)DeepSeek V4.1 FlashClaude Haiku 5.5: 65.4 Partial Credit accuracy % (v2.1), $1.64 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Haiku 5.5GLM 5.3: 65 Partial Credit accuracy % (v2.1), $1.03 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026) GLM 5.3Finance Agent Benchmark: score vs cost10 models. Frontier: DeepSeek V4.1 Flash, Gemini 3.8 Flash, Wonder Full, Gemini 4 Argon, Claude Opus 5.5.65666768$0.5$1$2$5$10EfficientInefficient↑ Best performance← Lowest costCost per test (USD)Claude Opus 5.5: 67.8 Partial Credit accuracy % (v2.1), $8.91 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Opus 5.5Gemini 4 Argon: 67.7 Partial Credit accuracy % (v2.1), $4.32 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 ArgonWonder Full: 67.6 Partial Credit accuracy % (v2.1), $2.68 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Wonder FullClaude Sonnet 5.5: 67.3 Partial Credit accuracy % (v2.1), $6.95 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Sonnet 5.5Gemini 4 Carbon: 67.2 Partial Credit accuracy % (v2.1), $7.29 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 4 CarbonGemini 3.8 Flash: 65.8 Partial Credit accuracy % (v2.1), $1.99 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Gemini 3.8 FlashWonder: 65.5 Partial Credit accuracy % (v2.1), $2.62 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)WonderDeepSeek V4.1 Flash: 65.4 Partial Credit accuracy % (v2.1), $0.37 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)DeepSeek V4.1 FlashClaude Haiku 5.5: 65.4 Partial Credit accuracy % (v2.1), $1.64 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026)Claude Haiku 5.5GLM 5.3: 65 Partial Credit accuracy % (v2.1), $1.03 per test (Vals.ai Finance Agent v2.1 leaderboard, Cost / Test, 29 Oct 2026) GLM 5.3

Finance Agent Benchmark v2.1

Partial Credit accuracy % (v2.1), higher is better. 10 models.

Updated 29 Oct 2026

  1. Claude Opus 5.567.8
  2. Gemini 4 Argon67.7
  3. Wonder Full67.6
  4. Claude Sonnet 5.567.3
  5. Gemini 4 Carbon67.2
  6. Gemini 3.8 Flash65.8
  7. Wonder65.5
  8. DeepSeek V4.1 Flash65.4
  9. Claude Haiku 5.565.4
  10. GLM 5.365

Finance Agent Benchmark: cost

cost per test item (USD), lower is better. 10 models.

Updated 29 Oct 2026

  1. DeepSeek V4.1 Flash0.370
  2. GLM 5.31.0
  3. Claude Haiku 5.51.6
  4. Gemini 3.8 Flash2.0
  5. Wonder2.6
  6. Wonder Full2.7
  7. Gemini 4 Argon4.3
  8. Claude Sonnet 5.57.0
  9. Gemini 4 Carbon7.3
  10. Claude Opus 5.58.9

Finance Agent Benchmark v2.1

Partial Credit accuracy % (v2.1), higher is better. 10 models.

Updated 29 Oct 2026

  1. 1Claude Opus 5.567.8
  2. 2Gemini 4 Argon67.7
  3. 3Wonder Full67.6
  4. 4Claude Sonnet 5.567.3
  5. 5Gemini 4 Carbon67.2
  6. 6Gemini 3.8 Flash65.8
  7. 7Wonder65.5
  8. 8DeepSeek V4.1 Flash65.4
  9. 9Claude Haiku 5.565.4
  10. 10GLM 5.365
Finance Agent Benchmark leaderboard (Partial Credit accuracy % (v2.1), higher is better)
1Claude Opus 5.5Anthropic67.8v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
2Gemini 4 ArgonGoogle67.7v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
3Wonder FullNot stated67.6v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
4Claude Sonnet 5.5Anthropic67.3v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
5Gemini 4 CarbonGoogle67.2v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
6Gemini 3.8 FlashGoogle65.8v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
7WonderNot stated65.5v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
8DeepSeek V4.1 FlashDeepSeek65.4v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
9Claude Haiku 5.5Anthropic65.4v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026
10GLM 5.3Zhipu AI65v2.1 test setVals.ai leaderboard
Official board, reported 29 Oct 2026 · checked 1 Nov 2026

Finance Agent Benchmark v2.1

Partial credit % on the 27 public questions (run by Moneda; not comparable, unranked), higher is better. 1 model.

Updated 9 Oct 2026

  1. Numa 1 (public set)80.2

Finance Agent Benchmark v2.1

Partial credit % on the 27 public questions (run by Moneda; not comparable, unranked), higher is better. 1 model.

Updated 9 Oct 2026

  1. 1Numa 1 (public set)80.2
Finance Agent Benchmark leaderboard (Partial credit % on the 27 public questions (run by Moneda; not comparable, unranked), higher is better)
1Numa 1 (public set)Moneda80.2Public set (27 questions), run by Moneda with its own grader (GPT-4o per rubric criterion); not a Vals result. Vals' other rows are its test set. One question is worth about 4 points.Moneda run, 2026-10-09
Run by Moneda, reported 9 Oct 2026

Last checked 1 Nov 2026. Something out of date or wrong? Send a correction.