Snorkel SWE-bench CLI
Code
Pass@1 on 200 repository engineering tasks spanning eleven languages, ten technical domains and eight task types. Harbor runs fail-to-pass and pass-to-pass tests in reproducible network-disabled Docker environments; credit requires all required tests and substantive changes spanning at least two files. Native effort and agent scaffold are not disclosed.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 16.6 |
| Claude Fable 5.1 | 14.5 |
| Claude Opus 5 | 14.0 |
| GPT-6 Astra | 12.4 |
| Gemini 3.8 Flash | 6.3 |
| Grok 4.7 | 6.1 |
| Grok 4.6 | 5.7 |
| Kimi K3 | 4.1 |
| GLM-5.3 | 3.5 |
| Muse Spark 1.3 | 2.2 |