Harvey LAB (Mercor run, undisclosed ~213-task cohort)
Legal
Rows of Mercor's Harvey LAB board that ran far fewer samples than its stated 1,749-task set: 545 to 639 each, where 639 is 213 tasks run three times, and their error margins also fit about 213 tasks. Mercor does not say which tasks these rows ran or whether every row ran the same ones, so the task count is left unknown and the rows are a diagnostic, never pooled with harvey-lab-mercor.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 4 Argon | 27.4 |
| Gemini 4 Carbon | 23.0 |
| Muse Wonder | 23.0 |
| Grok 4.7 | 19.5 |
| Muse Spark 1.2 | 19.4 |
| DeepSeek V4.1 Flash 1017 | 18.8 |
| Claude Fable 5.1 | 17.4 |
| Muse Spark 1.3 | 15.0 |
| GLM-5.3 | 14.6 |
| Mistral Large 4 Preview | 13.9 |
| MiniMax M3.1 Flash | 11.9 |
| Gemini 3.8 Flash | 11.4 |
| GLM-5.3 Flash | 11.1 |
| Claude Opus 5.5 | 10.8 |
| GPT-6 Astra | 9.2 |
| GPT-6.1 Sol | 8.8 |
| Mercury 3 | 7.8 |
| Beam 1.0 501B A23B | 7.7 |
| Claude Haiku 5.5 | 4.9 |
| GPT-6.1 Terra | 4.5 |
| Nemotron 3 Ultra 550B A55B (NVFP4) | 4.5 |
| GPT-6 Sol | 4.1 |
| GPT-5.6 Luna | 3.9 |
| GPT-6 Luna | 3.1 |
| GLM-5.1 | 2.7 |
| GPT-5.6 Terra | 2.3 |
| Gemini 3.5 Flash-Lite | 1.3 |