Vals AI LegalBench: leaderboard
Vals AI's run of LegalBench, a set of legal reasoning tasks written by legal professionals and grouped into six kinds of reasoning: spotting issues, recalling rules, applying them, drawing conclusions, interpreting text, and rhetorical understanding. Vals runs each model itself under one harness and breaks the results out by reasoning type. Higher accuracy is better. Source: arxiv.org/abs/2308.11462.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Estimated already saturated. 151 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (Max) | 88.56 | #16 (Claude Fable 5) |
| 2 | Claude Fable 5.1 (Max) | 88.51 | #9 (Claude Fable 5.1) |
| 3 | Gemini 4 Argon (High) | 88.3 | #8 (Gemini 4 Argon) |
| 4 | Muse Wonder Full (Max) | 87.42 | #10 (Muse Wonder Full) |
| 5 | Gemini 3.1 Pro (Preview) (High) | 87.4 | #58 (Gemini 3.1 Pro (Preview)) |
| 6 | Gemini 3.7 Flash (High) | 87.26 | #23 (Gemini 3.7 Flash) |
| 7 | Gemini 3 Pro (Preview) (High) | 87.03 | #66 (Gemini 3 Pro (Preview)) |
| 8 | Gemini 3.8 Flash (High) | 86.99 | #21 (Gemini 3.8 Flash) |
| 9 | GPT-5.6 Sol (Max) | 86.97 | #19 (GPT-5.6 Sol) |
| 10 | Claude Opus 5 (Max) | 86.97 | #12 (Claude Opus 5) |
| 11 | Gemini 3 Flash (Preview) (High) | 86.86 | #82 (Gemini 3 Flash (Preview)) |
| 12 | Gemini 3.6 Flash (High) | 86.7 | #51 (Gemini 3.6 Flash) |
| 13 | GPT-5.5 (xHigh) | 86.52 | #31 (GPT-5.5) |
| 14 | Grok 4.6 (High) | 86.31 | #26 (Grok 4.6) |
| 15 | Kimi K3 (Max) | 86.21 | #30 (Kimi K3) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-legalbench · How It Works · Data refreshed daily, snapshot 2026-10-20.