Harvey LAB-AA
Professional Work · 2026-07-07
Artificial Analysis' implementation of Harvey's Legal Agent Benchmark (LAB) evaluates AI agents on 120 private legal-work tasks spanning 24 practice areas. Agents use the Stirrup harness to read case documents in a sandbox and produce legal deliverables such as memos, disclosure schedules, and deposition summaries. The headline score is the all-pass rate: the percentage of tasks where every rubric criterion passes with no partial credit.
Top models (higher is better)
| Model | Score |
|---|---|
| Kimi K3 | 26.7 |
| Claude Fable 5 | 14.2 |
| Grok 4.5 | 13.3 |
| Muse Spark 1.1 | 8.3 |
| Opus 4.8 | 7.5 |
| GLM-5.2 | 7.5 |
| MiniMax M3 | 6.7 |
| Sonnet 5 | 5.0 |
| GPT-5.6 Luna | 5.0 |
| Sonnet 4.6 | 4.2 |
| GPT-5.5 | 4.2 |
| DeepSeek-V4-Pro | 3.3 |
| Nemotron 3 Ultra 550B A55B | 3.3 |
| GPT-5.6 Terra | 2.5 |
| Qwen3.6 27B | 2.5 |
| DeepSeek-V4-Flash | 1.7 |
| Gemini 3.5 Flash | 1.7 |
| GPT-5.6 Sol | 1.7 |
| Qwen3.7-Plus | 1.7 |
| Kimi K2.7 Code | 0.8 |
| Mistral Medium 3.5 | 0.8 |
| Haiku 4.5 | 0.0 |
| Gemini 3.1 Flash-Lite Preview | 0.0 |
| Gemini 3.1 Pro Preview | 0.0 |
| Gemma 4 31B IT | 0.0 |
| GPT-5.4 Mini | 0.0 |
| GPT-5.4 Nano | 0.0 |
| gpt-oss-120b | 0.0 |
| Grok 4.3 | 0.0 |
| Kimi K2.6 | 0.0 |
| MiMo-V2.5-Pro | 0.0 |
| Qwen3.5 397B A17B | 0.0 |
| Qwen3.7-Max | 0.0 |
| Step 3.7 Flash | 0.0 |