General intelligence

Vals Index

GDP-weighted agent performance across finance, coding, legal and tax work

What it measures

Vals Index is Vals AI’s composite measure of agent performance on real-world work, currently at version 2.2. It covers finance, coding, legal and tax tasks: financial research and Excel modeling; terminal work, app building and code migration; legal research and long-horizon legal work products; and research-grade U.S. corporate tax questions.

Think of it as a report card with differently weighted subjects: average each sector’s benchmark scores first, then combine the four sector averages. Finance has the greatest weight, so models with similar coding results can have different overall scores because of financial research and modeling.

Finance comprises Finance Agent v2.1 and the Excel Modeling Benchmark v1.1; coding comprises Terminal-Bench 4.0, Vibe Code Bench v1.1 and Code Migration v1.1; legal comprises Legal Research Bench and HLAB-Vals-1.1 (Harvey’s Legal Agent Benchmark); tax is Tax Agent Bench on its own. The weights use the sectors’ approximate shares of U.S.

GDP: finance and insurance 8.0, information 5.6, legal services 1.2, and accounting and tax preparation 0.5, totaling 15.3; normalized weights are approximately 52.3%, 36.6%, 7.8% and 3.3%.

Reading the scores

Higher scores mean stronger overall performance under this specific task mix and weighting. The formula is (8 × Finance + 5.6 × Coding + 1.2 × Legal + 0.5 × Tax) ÷ 15.3. For example, sector scores of 60, 80, 40 and 50 yield approximately 65.4%. This is neither a simple success rate across all tasks nor a percentage of jobs automated or GDP generated. This snapshot includes all 32 model results from the v2.2 leaderboard marked updated October 29, 2026, retaining published reasoning settings. V2.2 moves to Finance Agent v2.1, EMB v1.1, Code Migration v1.1 and HLAB-Vals-1.1, and switches Tax Agent Bench to all-pass accuracy. Components and weights change between versions, so their scores should not be compared directly. Code Migration uses a fixed subset of 50 CLI migration tasks plus all 10 COBOL tasks, weighted 75% / 25%; its full standalone leaderboard score cannot simply be substituted. We retain the published composite scores instead of recomputing them from other benchmarks on this site, and do not present source standard errors as 95% confidence intervals.

SourceVals AI

Explore the finance component: Finance Agent v2.1

Model performance

Checked November 1, 202632 results

Vals Index — Model performance. Select a column heading to sort.

Claude Sonnet 5.5 (max)Anthropic	70.9%¹
Gemini 4 Carbon (high)Google	70.4%
Claude Opus 5.5 (max)Anthropic	70.2%²
Gemini 4 Argon (high)Google	69.2%
Muse Wonder Full (max)Meta	66.8%
GPT-6 Astra (max)OpenAI	65.0%
Muse Wonder (max)Meta	64.9%
GPT-6.1 Sol (max)OpenAI	63.9%
Kimi K3.1 ProKimi	62.5%
DeepSeek V4.1 Flash 1017 (max)DeepSeek	61.3%
GPT-6.1 Terra (max)OpenAI	60.5%
GLM-5.3 (max)Z AI	59.7%
Qwen3.8 Max (0902)Alibaba	59.0%
Claude Haiku 5.5 (max)Anthropic	58.7%³
Grok 4.7 (xhigh)xAI	58.6%
MiMo V2.6 ProXiaomi	57.8%
Mistral Large 4 OSS (high)Mistral	56.4%
Gemini 3.8 Flash (high)Google	56.0%⁴
MiMo V2.6 FlashXiaomi	54.8%
Step 5 PreviewStepFun	54.8%
MiniMax-M3.1-Flash (max)MiniMax	53.8%
Hy4 PreviewTencent	53.2%
DeepSeek V4 Pro 0813 (max)DeepSeek	53.0%
Kimi K3Moonshot AI	53.0%
GPT-6 Luna (max)OpenAI	52.3%⁵
GLM-5.3 Flash (max)Z AI	50.9%
Muse Spark Open (xhigh)Meta	48.6%
Mercury 3 (high)Inception	43.7%⁶
Beam 1.0Reflection	43.4%
Nemotron 3.5 Super VL (high)NVIDIA	29.7%
Inkling (0.99)Thinking Machines	29.0%
Inkling Small (0.99)Thinking Machines	27.6%