SpatialBench: Can Agents Analyze Real-World Spatial Biology Data?
146 verifiable problems across five spatial technologies and seven task categories, from the scBench team. Each gives a data snapshot just before an analysis step plus a deterministic grader for recovering the key biological result. Accuracy is low (20–38% across families) and very sensitive to harness design — squarely in the spatial-omics wheelhouse. Seven frontier models evaluated; Claude Opus 4.5 leads at 38.4%.
Editor's note
146 verifiable problems across five spatial technologies and seven task categories, from the scBench team. Each gives a data snapshot just before an analysis step plus a deterministic grader for recovering the key biological result. Accuracy is low (20–38% across families) and very sensitive to harness design — squarely in the spatial-omics wheelhouse. Seven frontier models evaluated; Claude Opus 4.5 leads at 38.4%.
Results
| Model | Score | Variant | Provenance |
|---|---|---|---|
| Claude Opus 4.5 | 38.4% accuracy | default | paper |
| GPT-5.2 | 34.0% accuracy | default | paper |
| Claude Sonnet 4.5 | 28.3% accuracy | default | paper |
| GPT-5.1 | 27.4% accuracy | default | paper |
| Grok-4.1 | 24.7% accuracy | default | paper |
| Grok-4 | 22.8% accuracy | default | paper |
| Gemini 2.5 Pro | 20.1% accuracy | default | paper |
Sources
- paper SpatialBench (arXiv:2512.21907)