Harvey LAB (Mercor run, 1,749 public tasks)
Legal
Harvey's Legal Agent Benchmark as Mercor runs it: the 1,749 public LAB tasks across 24 practice areas and in-house contracting, in which an agent works through a client matter's files in a sandbox and delivers memos, redlines or filings. Mercor uses the Harvey LAB Agent harness and a Gemini 3 Flash Preview judge, a task passes only if every expert rubric criterion does, and Pass@1 averages three runs per task. The task set differs from the private held-out set behind hlab and hlab-v1-1, and the harvey-labs repository has since grown to 2,010 tasks, so this is its own benchmark.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 17.8 |
| Kimi K3 | 14.9 |
| Claude Fable 5 | 14.4 |
| DeepSeek V4 Pro 0813 | 11.4 |
| Grok 4.5 | 10.7 |
| Opus 4.8 | 9.4 |
| Gemini 3.7 Flash | 9.3 |
| Opus 4.7 | 8.7 |
| Sonnet 5 | 8.6 |
| MiniMax M3 | 5.3 |
| Gemini 3.6 Flash | 4.1 |
| GPT-5.5 | 4.1 |
| GPT-5.6 Sol Pro | 2.7 |
| Kimi K2.7 Code | 2.7 |
| MiniMax M2.7 | 1.9 |
| GPT-5.4 | 1.2 |
| Gemini 3.5 Flash | 0.7 |
| Gemini 3.1 Pro Preview | 0.1 |