Senior SWE-Bench
Code · 2026-06-30
Senior SWE-Bench is a Harbor-compatible software-engineering agent benchmark from Snorkel AI. Its v2026.06 suite contains 100 real-PR-derived tasks, with 50 public and 50 private. The displayed leaderboard uses model-specific scored coverage (93–95 tasks in the September 17 snapshot), so no single evaluation denominator is asserted. Score notes record each run's actual denominator. Tasteful solve rate (pass@1) requires runtime verifiers, validation, rubric, patch-bloat, practice, and relative-taste criteria.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 39.3 |
| GPT-6 Astra | 36.8 |
| Claude Fable 5.1 | 34.7 |
| Claude Fable 5 | 34.7 |
| Claude Opus 5 | 34.7 |
| GPT-5.6 Sol | 34.7 |
| Opus 4.8 | 30.5 |
| GPT-5.5 | 29.5 |
| Sonnet 5 | 28.0 |
| GPT-5.6 Terra | 27.4 |
| Grok 4.7 | 27.4 |
| Grok 4.6 | 26.3 |
| Grok 4.5 | 23.2 |
| DeepSeek V4.1 Flash | 22.7 |
| MiniMax M3 | 21.1 |
| Kimi K3 | 20.2 |
| Opus 4.7 | 18.1 |
| GLM-5.2 | 17.9 |
| GPT-5.4 | 16.8 |
| Gemini 3.7 Flash | 14.7 |
| Gemini 3.5 Flash | 6.3 |
| Gemini 3.1 Pro Preview | 2.1 |