LegalReasonBench: leaderboard
LegalReasonBench (micro1, 2026): 50 legal matters that practicing lawyers built end to end, each "a sequence of two to five assignments worked in one persistent sandbox" (122 assignments, 2,236 rubric criteria), from reading the case file and researching the law to court filings and memoranda. Unlike LegalBench, whose bar-exam style prompts use "a format that frontier models have since saturated", the client facts "stay in the uploaded case files rather than being paraphrased into the prompt". Protocol: the OpenCode 1.18.31 agent harness "with reasoning effort set to high for every model" in the report (shell with pdftotext, pandoc and Python, file tools, grep and glob, websearch and webfetch; web research is allowed in the 22 open-universe matters), a three-hour limit per assignment and three runs per matter; a strict majority of three LLM judges (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.8 Flash, each at medium reasoning effort) decides every criterion. Score: mean score (%), the mean over the 50 matters of each matter's mean reward over three runs; an assignment's reward is earned rubric points (penalties subtracted) over available positive points, clamped to [0, 1], and a matter's reward is the points-weighted total across its assignments. Rows labelled (High) are the report's runs; Kimi K3.1 Pro was added after the report and its run settings are not described; Gemini 4 Argon, which answered newer versions of 17 matters, is not listed. Higher is better. Source: https://www.micro1.ai/benchmark/legal-reason-bench
Metric: Mean score (%). Source: www.micro1.ai. Saturation forecast: Around 2030. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5.1 (High) | 50.6 | #10 (Claude Fable 5.1) |
| 2 | Claude Opus 5.5 (High) | 46.5 | #3 (Claude Opus 5.5) |
| 3 | Grok 4.7 (High) | 40.4 | #26 (Grok 4.7) |
| 4 | Muse Spark 1.3 (High) | 40.4 | #31 (Muse Spark 1.3) |
| 5 | Gemini 3.8 Flash (High) | 39.6 | #30 (Gemini 3.8 Flash) |
| 6 | GPT-6 Astra (High) | 39.5 | #6 (GPT-6 Astra) |
| 7 | Kimi K3.1 Pro | 36.4 | #17 |
| 8 | GPT-6 Sol (High) | 31.9 | #19 (GPT-6 Sol) |
No result here: #7 GPT-6.1 Sol, #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=legalreasonbench · How It Works · Data refreshed daily, snapshot 2026-11-05.