Benchmark
SWE-bench Science
What is SWE-bench Science?
Repository-level scientific software engineering benchmark from the OpenMOSS team (Fudan): 119 tasks across 98 GitHub repositories in 20 scientific domains…
- Released
- 2026-08-20
- Evaluates
- coding agent
- Openness
- unknown
- Importer
- Model reports
- Reported scores
- 8
Source provenance
- Original evidence https://arxiv.org/abs/2608.19799
SWE-bench Science paper
SWE-bench Science results and reported scores
Model reports
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Claude-Opus-5 (max) | Anthropic | 47.9 | 2026-09-01 | evidence |
| DeepSeek-V4-Flash-Max | DeepSeek | 23.53 | 2026-09-01 | evidence |
| DeepSeek-V4-Pro-Max | DeepSeek | 42.02 | 2026-09-01 | evidence |
| GLM-5.2 (max) | Zhipu AI | 31.93 | 2026-09-01 | evidence |
| GPT-5.6-sol (max) | OpenAI | 40.34 | 2026-09-01 | evidence |
| Kimi-K3 (max) | Moonshot AI | 35.29 | 2026-09-01 | evidence |
| Nex N2 | Nex | 24.37 | 2026-09-01 | evidence |
| Qwen3.5-397B-A17B | Alibaba | 14.29 | 2026-09-01 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite SWE-bench Science?
- SWE-bench Science results table OpenMOSS (Fudan University)