ScBench-Long V1 Clean Reruns
Science
Clean-v1 reruns on 22 long-horizon single-cell analysis evaluations. Formerly V2, now V1; the aggregate weight of 21 is not the task count.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Sonnet 5.5 | 31.8 |
| GPT-5.6 Sol | 29.6 |
| GPT-6 Sol | 28.4 |
| Claude Opus 5 | 27.3 |
| GPT-6 Astra | 27.3 |
| GPT-6.1 Sol | 26.7 |
| GPT-5.5 | 21.6 |
| Opus 4.8 | 21.0 |
| Gemini 3.8 Flash | 19.9 |
| Gemini 3.7 Flash | 18.8 |
| Grok 4.6 | 16.5 |
| Sonnet 5 | 15.9 |
| DeepSeek V4.1 Flash | 14.8 |
| Kimi K3 | 14.2 |
| Claude Opus 5.5 | 13.6 |
| GPT-5.6 Luna | 13.1 |
| Opus 4.7 | 11.4 |
| Grok 4.7 | 11.4 |
| GPT-6 Luna | 10.8 |
| Sonnet 4.6 | 10.2 |
| Nemotron 3.5 Lightning 30B A3B (precision unspecified) | 0.0 |