Proprietary

Vals Finance Index

Updated 10/21/2026

Agentic model performance on finance work, from analyst research and financial modeling to corporate tax research, weighted by each sub-sector's share of U.S. GDP.

As of October 21, 2026, Claude Opus 5.5 ranks first on the Vals Finance Index with 65.59%, followed by Gemini 4 Argon (64.89%) and Wonder Full (64.55%).

Vals Finance IndexGDP-weighted finance benchmark
ACCURACY

Vals Finance Index leaderboard

Rank Model Accuracy Cost / Test Input / Output Cost Input / Output Tokens Duration
1 Claude Opus 5.5 65.59% $11.67 $4 / $20 2.45B / 218.64M 45m00s
2 Gemini 4 Argon 64.89% $6.53 $4 / $20 2.02B / 60.38M 22m41s
3 Wonder Full 64.55% $3.14 $4 / $20 1.63B / 36.49M 20m06s
4 Claude Sonnet 5.5 64.53% $8.33 $2 / $10 4.72B / 286.30M 43m31s
5 Wonder 61.23% $3.37 $4 / $20 1.70B / 32.20M 10m12s
6 DeepSeek V4.1 Flash 1017 60.89% $0.42 $0.3 / $1.2 3.61B / 164.62M 23m02s
7 GLM 5.3 60.54% $2.21 $1.4 / $4.4 1.49B / 102.63M 35m44s
8 Gemini 3.8 Flash 60.46% $3.70 $1.5 / $7.5 4.33B / 55.08M 7m44s
9 Grok 4.7 58.46% $3.51 $2 / $6 1.55B / 48.97M 20m49s
10 Kimi K3 57.95% $2.58 $3 / $15 1.25B / 39.81M 22m24s
11 MiMo V2.6 Pro 57.85% $0.22 $0.435 / $0.87 1.28B / 41.27M 30m48s
12 Hy4 Preview 57.51% $0.68 $0.834 / $2.501 1.34B / 89.29M 32m15s
13 Claude Haiku 5.5 57.45% $1.39 $0.1 / $0.5 3.33B / 374.70M 33m31s
14 MiMo V2.6 Flash 57.15% $0.07 $0.14 / $0.28 1.45B / 42.78M 19m26s
15 DeepSeek V4.1 Flash 57.08% $0.25 $0.3 / $1.2 1.60B / 112.39M 11m37s
16 Mistral Large 4 56.90% $2.86 $1.36 / $4.18 1.70B / 73.58M 51m50s
17 MiniMax-M3.1-Flash 55.91% $0.25 $0.15 / $0.6 2.15B / 138.84M 26m12s
18 GLM 5.3 Flash 55.83% $0.31 $0.075 / $0.25 4.36B / 75.40M 33m49s
19 GPT-6.1 Sol 55.81% $1.36 $2 / $10 532.48M / 40.77M 19m16s
20 GPT-6 Astra 55.64% $6.48 $10 / $50 488.03M / 39.75M 16m18s
21 Step 5 Preview 55.43% $0.82 $1 / $2.7 1.98B / 51.31M 24m42s
22 GPT-6 Luna 53.05% $0.15 $0.1 / $0.5 1.20B / 97.59M 22m01s
23 Muse Spark Open 52.99% $3.00 $1.25 / $4.25 4.43B / 26.34M 41m04s
24 DeepSeek V4 Pro 0813 52.94% $0.88 $1.32 / $3.96 1.42B / 63.34M 24m00s
25 MiniMax-M3 48.37% $1.22 $0.6 / $2.4 2.56B / 29.44M 10m01s
26 Beam 1.0 46.50% N/A N/A 1.52B / 25.57M 7m29s
27 Inkling 38.25% $0.97 $1 / $4.05 1.65B / 43.31M 16m29s
28 Inkling Small 36.27% $0.18 $0.3 / $1.2 1.00B / 37.29M 10m24s
Vals Finance Index Over TimeBest score from each lab, by model release date
AccuracyTime
Vals Finance Index•Oct 21, 2026
Viewing 15 lab performance frontiers.
70Accuracy65605550454035
May '26Jun '26Jul '26Sep '26Oct '26

Key Takeaways

  • Even the leaders fail much of the hardest work. The top two are within the margin of error, and both still fail more than half of the tax questions and score about 75% on financial modeling.
  • No model leads every component. No single model leads every component: Claude Opus 5.5 leads Finance Agent v2.1 (67.85%) and Tax Agent Bench (45.57%); Claude Sonnet 5.5 leads EMB v1.1 (75.90%).
  • Close behind the leader for much less. Wonder Full and DeepSeek V4.1 Flash 1017 trail the leader by 1.0 and 4.7 points at 27% and 4% of its cost.
  • Rankings move a lot from the Vals Index. Gemini 3.8 Flash and Kimi K3 place far higher here than on the Vals Index; GPT-6 Astra and GPT-6.1 Sol place far lower.
  • Claude Sonnet 5.5 and Claude Opus 5.5 include tasks served by a fallback model after a refusal. The split icon on their rows shows the score without them.

Motivation

Finance is the largest sector of the Vals Index, but there its score is blended with coding and legal. The Finance Index looks at finance work on its own, tax included, using the same rule: each benchmark maps to an industry sub-sector, and each sub-sector is weighted by its share of U.S. GDP.

Methodology

Components

  • Securities & Investments (NAICS 523): Finance Agent v2.1, analyst research on public company filings, and the Excel Modeling Benchmark v1.1, building and editing financial models in Excel (LBO, DCF, M&A, comparables, operating models).
  • Tax (NAICS 5412): Tax Agent Bench, US corporate tax research questions answered with cited authority.

Each component column is that benchmark’s published standalone score. Tax Agent Bench contributes all-pass Accuracy, as it does on the Vals Index. A model is scored only when all three components have a result for it.

Weights

The Bureau of Economic Analysis splits finance and insurance into sub-sectors and files accounting and tax preparation under professional services. Each is weighted by its value added as a share of U.S. GDP, from BEA’s GDP by Industry accounts (annual 2025, September 2026 update; the finance and insurance total is also on FRED).

Sub-sector (NAICS)Share of GDPComponents
Securities, commodity contracts, and investments (523)1.6%Finance Agent v2.1, EMB v1.1
Accounting, tax preparation, bookkeeping, and payroll (5412)0.5%Tax Agent Bench

Formula

Each sub-sector is the average of its benchmarks, and the sub-sectors are combined by weight:

Securities = AVG(FinanceAgentV2_1, EMBV1_1)
Tax = TaxAgentBench
Finance_Index = (1.6 * Securities + 0.5 * Tax) / 2.1

Updates

10/21/2026

Released Vals Finance Index v1. Finance Agent v2.1 and EMB v1.1 form the Securities & Investments sub-sector (weight 1.6); Tax Agent Bench forms the Tax sub-sector (weight 0.5).