Harvey's Legal Agent Benchmark
Professional Work · 2026-05-06
Harvey's Legal Agent Benchmark held-out evaluation tests whether agents can complete realistic legal work using documents, spreadsheets, presentations, and file-system tools. The task score averages two judges' all-criteria-pass decisions, so individual tasks can receive 0, 0.5 or 1 credit. This is the Vals v1 leaderboard, archived at /benchmarks/hlab-v1 when HLAB-Vals-1.1 (hlab-v1-1, not comparable) replaced it in October 2026.
Top models (higher is better)
| Model | Score |
|---|---|
| Muse Spark 1.2 | 25.4 |
| Muse Spark 1.3 | 23.8 |
| Muse Wonder Full | 21.7 |
| Muse Spark 1.1 | 20.0 |
| Muse Wonder | 17.1 |
| Muse Spark Open | 16.7 |
| Grok 4.6 | 15.8 |
| Mistral Large 4 Preview | 15.8 |
| Claude Fable 5 | 13.3 |
| Grok 4.5 | 12.9 |
| Kimi K3 | 12.9 |
| Claude Opus 5 | 11.7 |
| MiMo-V2.6-Flash | 11.3 |
| Qwen3.8 27B | 11.3 |
| MiMo-V2.6-Pro | 10.8 |
| Qwen3.8-Max | 10.4 |
| Opus 4.8 | 10.4 |
| Gemini 3.8 Flash | 10.0 |
| DeepSeek V4.1 Flash 1017 | 9.2 |
| Ember-1 | 9.2 |
| Hy4 Preview | 9.2 |
| Step 5 Preview | 9.2 |
| Gemini 3.7 Flash | 8.8 |
| DeepSeek-V4-Flash-0731 | 8.3 |
| GLM-5.3 | 8.3 |
| DeepSeek V4 Pro 0813 | 7.5 |
| GLM-5.2 | 7.1 |
| Opus 4.7 | 6.7 |
| DeepSeek V4.1 Flash | 6.7 |
| GLM-5.3 Flash | 6.7 |
| Sonnet 5 | 5.8 |
| GPT-6.1 Sol | 5.4 |
| Sonnet 4.6 | 5.0 |
| MiniMax M3 | 4.2 |
| DeepSeek-V4-Pro | 3.8 |
| GPT-5.5 | 3.8 |
| Gemini 3.6 Flash | 3.3 |
| Claude Sonnet 5.5 | 2.9 |
| GPT-6 Luna | 2.9 |
| Gemini 3.5 Flash | 2.5 |