ScBench-Long V1 Verified
Science
ScBench-Long V1 verified merged 22-task cohort, distinct from the original 21-task release. Published per-model and per-harness pass rate on extended single-cell analysis workflows.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.6 Sol | 43.9 |
| GPT-6 Astra | 42.4 |
| Claude Opus 5 | 36.4 |
| GPT-5.6 Terra | 34.9 |
| GPT-5.6 Luna | 28.8 |
| DeepSeek V4.1 Flash | 27.3 |
| Grok 4.6 | 27.3 |
| Opus 4.8 | 25.8 |
| Grok 4.5 | 25.8 |
| Gemini 3.5 Flash | 24.2 |
| GPT-5.5 | 22.7 |
| Sonnet 5 | 19.7 |
| Kimi K3 | 16.7 |
| Opus 4.7 | 10.6 |