Harvey HLAB — Criteria Pass Rate
Professional Work · 2026-05-06
Harvey's Legal Agent Benchmark held-out v1 criteria pass rate (0–100), requiring both judges to pass each criterion. This strict criterion score is distinct from the judge-averaged task score and from v1.1's judge-averaged criteria definition.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.3 | 94.7 |
| Muse Spark 1.2 | 94.5 |
| Gemini 4 Argon | 94.0 |
| Muse Wonder Full | 93.5 |
| Muse Spark 1.1 | 92.9 |
| Muse Wonder | 92.7 |
| Grok 4.6 | 92.5 |
| Kimi K3 | 91.1 |
| MiMo-V2.6-Pro | 90.7 |
| Grok 4.5 | 90.5 |
| Grok 4.7 | 90.5 |
| Ember-1 | 90.2 |
| Gemini 3.8 Flash | 90.2 |
| Qwen3.8-Max | 90.0 |
| MiMo-V2.6-Flash | 89.9 |
| DeepSeek V4.1 Flash 1017 | 89.7 |
| GLM-5.3 | 89.7 |
| Hy4 Preview | 89.4 |
| Mistral Large 4 Preview | 89.3 |
| Step 5 Preview | 89.1 |
| DeepSeek-V4-Flash-0731 | 88.8 |
| Qwen3.8 27B | 88.8 |
| Gemini 3.7 Flash | 88.6 |
| DeepSeek V4 Pro 0813 | 88.1 |
| Opus 4.8 | 87.9 |
| GPT-6.1 Sol | 87.8 |
| Opus 4.7 | 87.8 |
| Claude Opus 5 | 87.7 |
| DeepSeek V4.1 Flash | 87.5 |
| GPT-6 Astra | 87.5 |
| Sonnet 4.6 | 86.7 |
| Sonnet 5 | 86.3 |
| MiniMax M3 | 86.3 |
| GLM-5.2 | 85.6 |
| DeepSeek-V4-Pro | 84.4 |
| Gemini 3.6 Flash | 84.4 |
| GPT-6 Sol | 83.5 |
| Qwen3.7-Max (source unspecified) | 83.5 |
| GPT-5.6 Luna | 83.4 |
| GPT-6 Luna | 83.4 |