SpatialBench-Long V1 Clean Reruns
Science
Clean-v1 reruns on 21 long-horizon spatial-transcriptomics evaluations. Formerly V2, now V1; the aggregate weight of 24 is not the task count.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6.1 Sol | 39.9 |
| GPT-6 Astra | 39.3 |
| GPT-5.6 Sol | 37.5 |
| Claude Sonnet 5.5 | 36.5 |
| Opus 4.8 | 33.3 |
| GPT-5.5 | 32.1 |
| GPT-6 Sol | 31.6 |
| Claude Opus 5 | 30.9 |
| Opus 4.7 | 29.8 |
| Gemini 3.7 Flash | 29.8 |
| Gemini 3.8 Flash | 28.6 |
| Grok 4.6 | 27.4 |
| DeepSeek V4.1 Flash | 26.2 |
| GPT-5.6 Luna | 25.0 |
| Sonnet 5 | 24.4 |
| Grok 4.7 | 24.4 |
| GPT-6 Luna | 23.2 |
| Sonnet 4.6 | 22.6 |
| Kimi K3 | 22.6 |
| Claude Opus 5.5 | 20.2 |
| Nemotron 3.5 Lightning 30B A3B (precision unspecified) | 4.8 |