Legal Research Bench
Whether an agent can research a real US legal question with case-law search, web search and document retrieval, then produce a supported answer — the multi-source work junior associates and paralegals do.
How it is run
Questions span eight practice areas of federal and state law and were written and peer-reviewed by practising lawyers, each with a rubric whose items are all required. Vals reports the strict all-pass rate as the headline and a weighted partial-credit figure beside it.
What it cannot tell you
Partial credit hides the failure mode: Opus 5 reaches 90.58% weighted but 55.29% all-pass, meaning models often get most of an answer right and miss one required element. Reconciling conflicting authority is the most reliable failure — every model loses 6 to 17 points on questions that need synthesis across jurisdictions, courts or regimes.
- Leader
- Muse Spark 1.3 Max55.29% · Vals AI's own list
- Top-ten spread
- 7.21%Leader minus tenth place, or last place in smaller fields.
- Median
- 32.94%of the 72 rows published
- No published result
- 9of 46 models on this site
Independent leaderboard · Vals AI
Top 72 of 72 runs by Vals AI, one harness.
| # | Model | Score |
|---|---|---|
| 1 | Muse Spark 1.3 Max | 55.29% |
| 1 | Claude Opus 5 | 55.29% |
| 1 | Claude Fable 5.1 | 55.29% |
| 4 | Gemini 4 Argon | 54.81% |
| 5 | Claude Opus 5.5 | 50.48% |
| 6 | Claude Fable 5 | 49.52% |
| 7 | GLM 5.3 | 49.04% |
| 8 | Grok 4.6 | 48.08% |
| 8 | GPT-5.6 Sol | 48.08% |
| 8 | Claude Sonnet 5.5 | 48.08% |
| 11 | Qwen3.8 Max | 47.60% |
| 12 | Grok 4.7 | 47.12% |
| 12 | MiMo V2.6 Pro | 47.12% |
| 14 | GLM 5.3 Flash | 45.19% |
| 14 | Hunyuan Hy4 Preview | 45.19% |
| 16 | Ember 1 | 44.71% |
| 17 | Kimi K3 | 44.23% |
| 18 | Muse Spark 1.2 | 43.75% |
| 18 | Claude Opus 4.8 | 43.75% |
| 20 | Claude Sonnet 5 | 41.83% |
| 21 | DeepSeek V4.1 Flash | 41.35% |
| 21 | GPT-5.6 Terra | 41.35% |
| 23 | Muse Spark 1 3 | 40.87% |
| 23 | DeepSeek V4 Pro | 40.87% |
| 25 | GPT-5.5 | 40.38% |
| 26 | GPT-6 Astra | 39.42% |
| 27 | Gemini 3.8 Flash | 38.94% |
| 28 | Claude Opus 4.7 | 38.46% |
| 28 | Claude Sonnet 4 6 | 38.46% |
| 28 | GPT-6.1 Sol | 38.46% |
| 31 | Muse Spark 1 1 | 37.98% |
| 31 | MiMo-V2.6-Flash | 37.98% |
| 31 | Grok 4.5 | 37.98% |
| 34 | GPT-5.6 Luna | 36.54% |
| 35 | Qwen3.8 27b | 36.06% |
| 36 | Gemini 3.7 Flash | 34.62% |
| 37 | Glm 5.2 | 31.25% |
| 38 | Gemini 3.5 Flash | 30.77% |
| 39 | Deepseek V4 Flash 0731 | 30.29% |
| 39 | GPT-6 Luna | 30.29% |
| 41 | MiniMax M3 | 29.81% |
| 42 | GPT-6 Sol | 28.85% |
| 43 | Inkling | 28.36% |
| 44 | Glm 5.1 | 27.89% |
| 45 | Inkling Small | 25.48% |
| 45 | Qwen3.7 Max | 25.48% |
| 47 | Gemini 3.6 Flash | 25.00% |
| 48 | Deepseek V4 Pro | 23.08% |
| 49 | Gemini 3.1 Pro | 20.67% |
| 50 | Gemini 3 Flash Preview | 18.27% |
| 51 | Qwen3.7 Plus | 16.35% |
| 52 | Mimo V2.5 Pro | 15.87% |
| 52 | Kimi K2.5 Thinking | 15.87% |
| 52 | Kimi K2.6 | 15.87% |
| 55 | Grok 4.3 | 15.38% |
| 55 | Ling 3.0 Flash Af Rc3 | 15.38% |
| 55 | Nemotron 3 Ultra 550b A55b | 15.38% |
| 58 | Qwen3.6 Plus | 14.90% |
| 59 | Gemini 3.5 Flash Lite | 13.94% |
| 59 | Grok 4.20 0309 Reasoning | 13.94% |
| 61 | Gpt 5.4 Mini 2026 03 17 | 12.50% |
| 62 | Claude Haiku 4 5 20251001 Thinking | 10.58% |
| 62 | MiniMax M2.7 | 10.58% |
| 64 | Mimo V2.5 | 9.13% |
| 64 | Mistral Medium 3.5 | 9.13% |
| 66 | Gpt 5.4 Nano 2026 03 17 | 6.25% |
| 67 | Mercury 2.5 | 4.33% |
| 68 | Gemini 3.1 Flash Lite Preview | 3.37% |
| 69 | Nemotron Lightning 3p5 30b A3b | 2.88% |
| 70 | Laguna M.1 | 2.40% |
| 71 | Laguna Xs.2 | 0.96% |
| 72 | Ling 3.0 Flash 2607 | 0.00% |
No published result tracked
9 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.
Read this with
Benchmarks in the same category, which usually disagree with this one on purpose.
Sources
Every page used for this benchmark, read on 17 October 2026.
- Benchmark site or the page that publishes itwww.vals.ai/benchmarks/legal_research
- Independent resultswww.vals.ai/benchmarks/legal_research