Benchmarks · Vals AI
LegalBench (Vals leaderboard)
Note: Vals has no benchmark aimed at Dutch or EU law. Measures legal reasoning on short, well-delimited tasks (open dataset by Guha et al.; Vals runs every model with the same settings). Split into issue, rule, conclusion, interpretation and rhetoric tasks. The top ten sit between 87.0 and 88.6 percent (as at 13 October 2026), so the ranking separates little. The original dataset (https://github.com/HazyResearch/legalbench) carries a licence per task; its maintainers ask users to follow the dataset creator's licence.
- Models measured
- 113
- Top
- 88.56 %Claude Fable 5within the standard error of ±0.33
- Version
- 1
- Measured
- 13 October 2026
01
Ranking
At the top: Top 15
Vals AI Legal Research Bench, all-pass [1]. Hatched bars are rejected by the compliance gate.
- 1Claude Fable 588.56%
- 2Claude Fable 5.188.51%
- 3Gemini 4 Argon88.30%
- 4Wonder Full87.42%
- 5Gemini 3.1 Pro Preview (02/26)87.40%
- 6Gemini 3.7 Flash87.26%
- 7Gemini 3 Pro (11/25)87.03%
- 8Gemini 3.8 Flash86.99%
- 9Claude Opus 586.97%
- 10GPT-5.6 Sol86.97%
- 11Gemini 3 Flash (12/25)86.86%
- 12Gemini 3.6 Flash86.70%
- 13GPT 5.586.52%
- 14Grok 4.686.31%
- 15GLM Legal86.23%
Source: Vals AI (LegalBench (Vals leaderboard)) · Licence: CC BY 4.0 · Adaptation: Prudai selected the models, translated the explanations, added the gate verdict and the PrudentScore and rounded figures to two decimals.
| # Model | Score | Cost/task | Gate | PrudentScore |
|---|---|---|---|---|
| 1Claude Fable 5Anthropic | 88.56%± 0.33 | - | Rejected | no score |
| 2Claude Fable 5.1Anthropic | 88.51%± 0.42 | - | Rejected | no score |
| 3Gemini 4 ArgonGoogle | 88.30%± 0.37 | - | Passes | 79 |
| 4Wonder FullMeta | 87.42%± 0.41 | - | Rejected | no score |
| 5Gemini 3.1 Pro Preview (02/26)Google | 87.40%± 0.33 | - | Rejected | no score |
| 6Gemini 3.7 FlashGoogle | 87.26%± 0.42 | - | Passes | 58 |
| 7Gemini 3 Pro (11/25)Google | 87.03%± 0.37 | - | Rejected | no score |
| 8Gemini 3.8 FlashGoogle | 86.99%± 0.43 | - | Passes | 62 |
| 9Claude Opus 5Anthropic | 86.97%± 0.42 | - | Passes | 72 |
| 10GPT-5.6 SolOpenAI | 86.97%± 0.41 | - | Passes | 57 |
| 11Gemini 3 Flash (12/25)Google | 86.86%± 0.36 | - | Rejected | no score |
| 12Gemini 3.6 FlashGoogle | 86.70%± 0.41 | - | Passes | 46 |
| 13GPT 5.5OpenAI | 86.52%± 0.41 | - | Passes | 51 |
| 14Grok 4.6xAI (SpaceXAI) | 86.31%± 0.42 | - | Rejected | no score |
| 15GLM LegalZ.ai | 86.23%± 0.41 | - | Rejected | no score |
| 16Kimi K3Moonshot AI | 86.21%± 0.41 | - | Rejected | no score |
| 17GPT 5.4 (xhigh)OpenAI | 86.04%± 0.41 | - | Passes | 0 |
| 18GPT 5OpenAI | 86.02%± 0.38 | - | Passes | no score |
| 19Grok 4.5xAI (SpaceXAI) | 85.97%± 0.41 | - | Rejected | no score |
| 20GPT 5.1OpenAI | 85.68%± 0.40 | - | Passes | no score |
| 21MiniMax-M3MiniMax | 85.42%± 0.44 | - | Rejected | no score |
| 22Claude Opus 4.6 (Thinking)Anthropic | 85.30%± 0.37 | - | Passes | no score |
| 23Muse Spark 1.2Meta | 85.26%± 0.45 | - | Rejected | no score |
| 24Claude Opus 4.7Anthropic | 85.25%± 0.45 | - | Passes | 54 |
| 25GPT-5.6 TerraOpenAI | 85.11%± 0.45 | - | Passes | 53 |
| 26Qwen 3.5 PlusAlibaba | 85.10%± 0.41 | - | Not yet assessed | no score |
| 27Muse Spark 1.1Meta | 84.98%± 0.42 | - | Rejected | no score |
| 28Qwen 3.7 MaxAlibaba | 84.91%± 0.41 | - | Not yet assessed | no score |
| 29GLM 5.3Z.ai | 84.84%± 0.40 | - | Rejected | no score |
| 30Kimi K2.6Moonshot AI | 84.74%± 0.45 | - | Rejected | no score |
| 31Claude Opus 4.5 (Thinking)Anthropic | 84.60%± 0.39 | - | Passes | no score |
| 32Grok 4.3xAI (SpaceXAI) | 84.46%± 0.42 | - | Rejected | no score |
| 33GLM 5.1Z.ai | 84.39%± 0.43 | - | Rejected | no score |
| 34Grok 4.7xAI (SpaceXAI) | 84.39%± 0.46 | - | Rejected | no score |
| 35Qwen 3.5 FlashAlibaba | 84.28%± 0.41 | - | Not yet assessed | no score |
| 36Qwen 3.6 PlusAlibaba | 84.23%± 0.42 | - | Not yet assessed | no score |
| 37Muse SparkMeta | 84.22%± 0.42 | - | Rejected | no score |
| 38Claude Sonnet 4.5 (Thinking)Anthropic | 84.08%± 0.45 | - | Passes | no score |
| 39=Gemini 3.5 Flash LiteGoogle | 84.07%± 0.44 | - | Passes | 38 |
| 39=GLM 5.2Z.ai | 84.07%± 0.45 | - | Rejected | no score |
| 41GLM 5Z.ai | 84.06%± 0.41 | - | Passes | no score |
| 42GPT-5.6 LunaOpenAI | 84.03%± 0.42 | - | Passes | 56 |
| 43MiniMax-M2.7MiniMax | 83.98%± 0.41 | - | Rejected | no score |
| 44GLM 5.3 FlashZ.ai | 83.93%± 0.43 | - | Rejected | no score |
| 45Claude Sonnet 5Anthropic | 83.92%± 0.46 | - | Passes | 59 |
| 46Gemini 3.1 Flash Lite PreviewGoogle | 83.76%± 0.41 | - | Not yet assessed | no score |
| 47o3OpenAI | 83.76%± 0.42 | - | Passes | no score |
| 48Hy4 PreviewTencent | 83.76%± 0.41 | - | Rejected | no score |
| 49Qwen 3.8 MaxAlibaba | 83.61%± 0.42 | - | Not yet assessed | no score |
| 50Gemini 3.5 FlashGoogle | 83.60%± 0.85 | - | Passes | 47 |
| 51Claude Opus 4.8Anthropic | 83.57%± 0.47 | - | Passes | 62 |
| 52Claude Opus 4.1 (Nonthinking)Anthropic | 83.46%± 0.45 | - | Rejected | no score |
| 53GLM 4.7Z.ai | 83.36%± 0.47 | - | Passes | no score |
| 54DeepSeek V4.1 FlashDeepSeek | 83.28%± 0.46 | - | Rejected | no score |
| 55Grok 4xAI (SpaceXAI) | 83.19%± 0.47 | - | Rejected | no score |
| 56GPT 4.1OpenAI | 83.10%± 0.43 | - | Passes | no score |
| 57Claude Opus 4 (Nonthinking)Anthropic | 83.07%± 0.44 | - | Rejected | no score |
| 58Mercury 2.5Inception Labs | 83.06%± 0.48 | - | Not yet assessed | no score |
| 59Inkling SmallThinking Machines | 82.99%± 0.46 | - | Rejected | no score |
| 60Claude Sonnet 4 (Nonthinking)Anthropic | 82.95%± 0.41 | - | Rejected | no score |
| 61InklingThinking Machines | 82.95%± 0.50 | - | Rejected | no score |
| 62Claude Opus 4.5 (Nonthinking)Anthropic | 82.84%± 0.43 | - | Passes | no score |
| 63GPT 5.2OpenAI | 82.76%± 0.40 | - | Passes | no score |
| 64Gemini 2.5 Flash Preview (9/25) (Thinking)Google | 82.63%± 0.41 | - | Rejected | no score |
| 65Gemini 2.5 Flash Preview (9/25) (Nonthinking)Google | 82.49%± 0.43 | - | Rejected | no score |
| 66Grok 4.1 Fast (Reasoning)xAI (SpaceXAI) | 82.45%± 0.46 | - | Rejected | no score |
| 67Qwen 3.8 27BAlibaba | 82.43%± 0.47 | - | Rejected | no score |
| 68DeepSeek V4 Pro (0813)DeepSeek | 82.36%± 0.44 | - | Rejected | no score |
| 69GPT 4o (2024-11-20)OpenAI | 82.21%± 0.46 | - | Passes | no score |
| 70Claude Sonnet 4.6Anthropic | 82.12%± 0.48 | - | Passes | 55 |
| 71Nemotron 3 UltraNVIDIA | 82.07%± 0.46 | - | Rejected | no score |
| 72Claude Sonnet 4 (Thinking)Anthropic | 82.06%± 0.46 | - | Rejected | no score |
| 73Gemini 2.5 Flash Lite (9/25) (Thinking)Google | 82.04%± 0.44 | - | Rejected | no score |
| 74Qwen 3 MaxAlibaba | 81.86%± 0.40 | - | Not yet assessed | no score |
| 75GPT 5 MiniOpenAI | 81.77%± 0.45 | - | Passes | no score |
| 76Kimi K2 InstructMoonshot AI | 81.45%± 0.44 | - | Rejected | no score |
| 77Claude Haiku 4.5 (Thinking)Anthropic | 81.24%± 0.49 | - | Passes | 32 |
| 78Grok 4 Fast (Reasoning)xAI (SpaceXAI) | 80.60%± 0.45 | - | Rejected | no score |
| 79Qwen 3 Max PreviewAlibaba | 80.33%± 0.40 | - | Not yet assessed | no score |
| 80DeepSeek V4 Pro (2026-04-23)DeepSeek | 80.32%± 0.47 | - | Rejected | no score |
| 81Kimi K2 ThinkingMoonshot AI | 80.20%± 0.46 | - | Rejected | no score |
| 82MiniMax-M2.5MiniMax | 79.96%± 0.47 | - | Passes | no score |
| 83Command ACohere | 79.70%± 0.42 | - | Rejected | no score |
| 84Ling 3.0 FlashAnt Group | 79.69%± 0.52 | - | Rejected | no score |
| 85GLM 4.6Z.ai | 79.61%± 0.52 | - | Rejected | no score |
| 86o4 MiniOpenAI | 79.19%± 0.50 | - | Passes | no score |
| 87Mistral Large 3Mistral AI | 79.14%± 0.46 | - | Passes | no score |
| 88Grok 4.1 Fast Non-ReasoningxAI (SpaceXAI) | 79.11%± 0.46 | - | Rejected | no score |
| 89Gemini 2.5 Flash Lite (9/25) (Nonthinking)Google | 79.01%± 0.44 | - | Rejected | no score |
| 90MiMo V2.5Xiaomi | 78.89%± 0.51 | - | Rejected | no score |
| 91Grok 4 Fast (Non-Reasoning)xAI (SpaceXAI) | 78.40%± 0.47 | - | Rejected | no score |
| 92GPT 4.1 MiniOpenAI | 78.04%± 0.42 | - | Passes | no score |
| 93GPT 5.4 NanoOpenAI | 77.92%± 0.43 | - | Passes | 29 |
| 94Llama 4 MaverickMeta | 77.81%± 0.42 | - | Rejected | no score |
| 95Grok 4.20 (Reasoning)xAI (SpaceXAI) | 77.74%± 0.48 | - | Rejected | no score |
| 96DeepSeek V4 Flash 0731DeepSeek | 77.71%± 0.50 | - | Rejected | no score |
| 97MiMo V2.5 ProXiaomi | 77.13%± 0.55 | - | Rejected | no score |
| 98DeepSeek V3.2 (Thinking)DeepSeek | 76.08%± 0.49 | - | Passes | no score |
| 99GPT OSS 120BOpenAI | 75.94%± 0.52 | - | Passes | no score |
| 100GLM 4.5Z.ai | 75.63%± 0.51 | - | Rejected | no score |
| 101MiniMax-M2.1MiniMax | 75.45%± 0.54 | - | Passes | no score |
| 102Laguna M.1Poolside | 75.14%± 0.59 | - | Rejected | no score |
| 103Llama 4 ScoutMeta | 72.04%± 0.48 | - | Rejected | no score |
| 104Laguna XS.2Poolside | 71.03%± 0.56 | - | Rejected | no score |
| 105GPT OSS 20BOpenAI | 70.85%± 0.61 | - | Passes | no score |
| 106Mistral Small 3.1 (03/2025)Mistral AI | 69.16%± 0.51 | - | Passes | no score |
| 107DeepSeek V3.2 (Nonthinking)DeepSeek | 65.96%± 0.42 | - | Passes | no score |
| 108Mistral Medium 3.1 (05/2025)Mistral AI | 61.98%± 0.47 | - | Passes | no score |
| 109Command A+Cohere | 61.42%± 0.63 | - | Rejected | no score |
| 110GPT 4.1 NanoOpenAI | 61.06%± 0.59 | - | Passes | no score |
| 111Magistral Medium 1.2 (09/2025)Mistral AI | 54.77%± 0.56 | - | Not yet assessed | no score |
| 112GPT 5 NanoOpenAI | 50.13%± 0.47 | - | Passes | no score |
| 113Magistral Small 1.2 (09/2025)Mistral AI | 40.02%± 0.47 | - | Passes | no score |
Source: Vals AI (LegalBench (Vals leaderboard)) · Licence: CC BY 4.0 · Adaptation: Prudai selected the models, translated the explanations, added the gate verdict and the PrudentScore and rounded figures to two decimals.
Vals AI · Version 1 · Measured 13 October 2026 · Go to the source. The PrudentScore is a composition by Prudai B.V. based on publicly published figures and does not represent the view or endorsement of Vals AI, Harvey or Artificial Analysis. Figures from Artificial Analysis appear in running text only, with attribution and a link, as that party's terms require [9].
Other benchmarks
- Legal Research Bench
- Harvey LAB (Vals implementation)
- CaseLaw v2
- TaxEval v2
- Tax Agent Bench
- CorpFin v2
- Immigration Bench
- Access to Justice
- Public Benefits Bench
- Vals Legal Index
- Vals Index
- GPQA Diamond (Vals-run)
- MMLU Pro (Vals-run)
- Artificial Analysis Intelligence Index
Claude and Anthropic, GPT and OpenAI, Gemini and Google, Grok and xAI, Kimi and Moonshot AI, Muse and Meta, Mistral, DeepSeek, GLM and Z.ai, Harvey, Vals AI and Artificial Analysis are trademarks of their respective owners. They are named for identification only and this implies no partnership, sponsorship or endorsement.