SWE-rebench: leaderboard
Each task is a GitHub issue mined continuously from active open source repositories and scored by date window, with runs flagged when issues predate a model's release. The SWE-rebench team runs every model in one fixed ReAct harness with a 128k context and terminal, judging patches by the fixing pull request's tests. Models are run five times and the board shows the mean resolved rate in percent, higher is better.
Metric: Resolved (%). Source: swe-rebench.com. Saturation forecast: Around January 2027. 97 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Fable 5 (High) | 64.5 | #16 (Claude Fable 5) |
| 2 | Grok 4.5 (High) | 63.78 | #48 (Grok 4.5) |
| 3 | Claude Opus 5 (High) | 63.42 | #12 (Claude Opus 5) |
| 4 | GLM-5.1 | 62.67 | #97 |
| 5 | GPT-5.6 Sol (Medium) | 62.34 | #19 (GPT-5.6 Sol) |
| 6 | Qwen 3.5 397B A17B | 59.94 | #146 |
| 7 | Step 3.5 Flash | 59.58 | #270 |
| 8 | GPT-5.2 (2025-12-11) (Medium) | 59.25 | |
| 9 | Qwen 3.5 27B | 58.95 | #187 |
| 10 | GPT-5.3 Codex (xHigh) | 58.6 | #69 (GPT-5.3 Codex) |
| 11 | GPT-5.3 Codex | 58.21 | #69 |
| 12 | GPT-5.4 (2026-03-05) (Medium) | 57.6 | #59 (GPT-5.4 (2026-03-05)) |
| 13 | GLM-5.2 (High) | 57.01 | #73 (GLM-5.2) |
| 14 | Claude Sonnet 5 (High) | 56.76 | #57 (Claude Sonnet 5) |
| 15 | GPT-5.2 (2025-12-11) (xHigh) | 56.57 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swe-rebench · How It Works · Data refreshed daily, snapshot 2026-10-20.