ScBench-Long
Science · 2026-06-25
Long-horizon single-cell analysis benchmark on real single-cell RNA-seq datasets, requiring extended multi-step agentic workflows such as quality control, clustering, cell-type annotation, and downstream analysis. Scores average three independent trials for each of the 21 distinct tasks, for 63 task runs.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 41.3 |
| GPT-5.6 Sol | 38.1 |
| Opus 4.8 | 25.4 |
| GPT-5.6 Terra | 23.8 |
| Gemini 3.5 Flash | 22.2 |
| GPT-5.6 Luna | 19.1 |
| Gemini 3.1 Pro Preview | 17.5 |
| GPT-5.5 | 15.9 |
| Sonnet 5 | 14.3 |
| Grok 4.5 | 12.7 |
| Kimi K3 | 12.7 |
| Opus 4.7 | 11.1 |
| Opus 4.6 | 9.5 |
| GLM-5.2 | 9.5 |
| GPT-5.4 | 9.5 |
| Sonnet 4.6 | 7.9 |
| Grok 4.20 | 6.3 |
| Kimi K2.6 | 4.8 |
| Grok 4.3 | 1.6 |