ChainSWE (Oracle): leaderboard
ChainSWE (2026) links 304 real Python issues from six SWE-bench-family datasets into 100 chronological chains of dependent bug fixes in 54 repositories and runs each model in the SWE-agent-equivalent SWE-Edit baseline scaffold at medium thinking effort. In this oracle mode every bug starts from its chain with the earlier gold fixes applied; score is the per-bug resolution rate in percent, higher is better.
Metric: Per-bug resolution rate (%; share of the 304 bugs in 100 chronological bug chains from six SWE-bench-family datasets whose tests pass, each bug starts from its chain with the earlier gold fixes applied; the Baseline configuration of the SWE-Edit scaffold, which the paper states is identical to the original SWE-agent; 100 turns and 30 minutes per bug, thinking effort medium). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 69.1 | #49 |
| 2 | GPT-5.4 Nano (Medium) | 64.8 | #635 (GPT-5.4 Nano) |
| 3 | Claude Opus 4.7 (Medium) | 64.5 | #79 (Claude Opus 4.7) |
| 4 | Gemini 3.1 Pro (Preview) (Medium) | 61.8 | #94 (Gemini 3.1 Pro (Preview)) |
| 5 | GPT-5.4 Mini (Medium) | 61.8 | #389 (GPT-5.4 Mini) |
| 6 | DeepSeek V4 Pro (Medium) | 42.8 | #188 (DeepSeek V4 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=chainswe-oracle · How It Works · Data refreshed daily, snapshot 2026-11-13.