LegalScope - Constraint Extraction: leaderboard
LegalScope: 28 model configurations; automatic grading on 861 exam and 276 case prompts, human validation on 80 exam and 10 case prompts with two lawyers. Scores scaled 0-100. August generation logs explicitly distinguish non-thinking and Low/Medium/High settings; unspecified older settings remain unspecified. Later provider defaults do not rewrite the experiment.. Score (%).
Metric: Score (%). Source: raw.githubusercontent.com. Saturation forecast: Around January 2027. 28 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (High) | 98.7 | #29 (GPT-5.5) |
| 2 | GPT-5.5 (Medium) | 98.6 | #29 (GPT-5.5) |
| 3 | GPT-5.4 Mini | 90.2 | #212 |
| 4 | Grok 4.5 | 89.7 | #49 |
| 5 | GLM-5.3 (Low) | 87 | #36 (GLM-5.3) |
| 6 | DeepSeek V4 Pro (Non-reasoning) | 86.1 | #106 (DeepSeek V4 Pro) |
| 7 | Grok 4.3 | 85.3 | #190 |
| 8 | Gemini 2.5 Flash | 84.6 | #242 |
| 9 | Qwen 3 Max (Thinking) | 82.5 | #210 (Qwen 3 Max) |
| 10 | Qwen 3 Max | 81.8 | #210 |
| 11 | Qwen 3.8 27B (Non-reasoning) | 80.3 | #118 (Qwen 3.8 27B) |
| 12 | Command A | 80.2 | #503 |
| 13 | DeepSeek V3.2 | 79.3 | #204 |
| 14 | DeepSeek R1 | 79.3 | #251 |
| 15 | Grok 4.20 | 76.3 | #172 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=legalscope-constraint-extraction · How It Works · Data refreshed daily, snapshot 2026-10-23.