Vals AI Legal Research Bench: leaderboard
Agentic legal research on US federal and state law across eight practice areas: the model researches a legal question with web search, page parsing, document retrieval and CourtListener case law search. GPT 5.4 judges the answer against a rubric written and peer reviewed by practicing lawyers. The number is the share of the 208 private test tasks whose answer satisfies every rubric item, and higher is better.
Metric: Accuracy (%). Source: www.vals.ai. Saturation forecast: Around May 2027. 80 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5 | 55.29 | #12 |
| 2 | Claude Fable 5.1 | 55.29 | #9 |
| 3 | Muse Spark 1.3 (Max) | 55.29 | #24 (Muse Spark 1.3) |
| 4 | Gemini 4 Argon | 54.81 | #8 |
| 5 | Muse Wonder Full | 53.37 | #10 |
| 6 | Claude Opus 5.5 | 50.48 | #3 |
| 7 | Muse Wonder | 50.48 | #15 |
| 8 | Claude Fable 5 | 49.52 | #16 |
| 9 | GLM-5.3 | 49.04 | #39 |
| 10 | GPT-5.6 Sol | 48.08 | #19 |
| 11 | Grok 4.6 | 48.08 | #26 |
| 12 | Claude Sonnet 5.5 | 48.08 | #11 |
| 13 | Qwen 3.8 Max | 47.6 | #35 |
| 14 | MiMo-V2.6-Pro | 47.12 | #42 |
| 15 | Grok 4.7 | 47.12 | #25 |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-legal-research-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.