Harvey LAB-AA v1.0 (criterion pass rate)
Professional Work · 2026-07-07
Artificial Analysis' original Harvey LAB-AA v1.0 protocol evaluates agents on 120 private legal-work tasks spanning 24 practice areas using the Stirrup harness. The archived harveyLab field reports criterion pass rate, the share of rubric criteria satisfied under the original single-judge protocol. These scores are not comparable to v1.1, which revises the tasks and judges and gates its headline score on material hallucinations.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.3 | 95.5 |
| Kimi K3 | 94.6 |
| Qwen3.8 Max (0902) | 93.6 |
| Claude Fable 5 | 93.6 |
| Claude Opus 5 | 93.5 |
| Step 5 Preview | 93.4 |
| Claude Fable 5.1 | 93.3 |
| Muse Spark 1.1 | 93.1 |
| Claude Sonnet 5.5 | 93.1 |
| Grok 4.5 | 92.4 |
| Claude Opus 5.5 | 91.2 |
| Opus 4.8 | 91.1 |
| GLM-5.2 | 91.0 |
| Gemini 3.7 Flash | 90.7 |
| Sonnet 5 | 90.1 |
| Claude Haiku 5.5 | 89.9 |
| MiniMax M3 | 88.4 |
| GPT-5.6 Luna | 87.9 |
| GPT-5.6 Sol | 87.2 |
| GPT-5.5 | 86.3 |
| Sonnet 4.6 | 86.0 |
| GPT-5.6 Terra | 85.2 |
| Gemini 3.6 Flash | 85.1 |
| Kimi K2.7 Code | 85.0 |
| DeepSeek-V4-Pro | 84.4 |
| Kimi K2.6 | 84.1 |
| Qwen3.7-Max (source unspecified) | 83.4 |
| Qwen3.6 27B | 82.3 |
| Gemini 3.5 Flash | 82.1 |
| Qwen3.7-Plus (source unspecified) | 81.8 |
| Nemotron 3 Ultra 550B A55B | 81.7 |
| DeepSeek-V4-Flash | 81.3 |
| MiMo-V2.5-Pro | 73.3 |
| Step 3.7 Flash | 72.7 |
| Qwen3.5 397B A17B | 72.4 |
| Mistral Medium 3.5 | 69.1 |
| Grok 4.3 | 68.4 |
| Haiku 4.5 | 61.1 |
| GPT-5.4 Mini | 60.7 |
| Gemini 3.1 Pro Preview | 58.9 |