SWE-rebench
About
Evaluates LLM coding agents on real-world GitHub issues sourced after each model's training cutoff, preventing benchmark contamination. Uses standardized ReAct scaffolding with 128K token context; each model is run five times per problem and the best Pass@1 resolved rate is reported.
How to read this benchmark
This benchmark uses a scoring system where higher is better. Scores are useful for directional filtering and model shortlisting, not for universal quality ranking.
Interpretation checklist: prefer benchmarks that are closest to your workload style, then validate against the linked model pages for pricing, context window, and provider availability.
Trust when:
- There is a fresh timestamped snapshot (or multiple snapshots) for this benchmark.
- Model list covers the same version family you can actually deploy today.
- Top candidates overlap with your required routing and feature requirements.
Don't trust when:
- There is only one benchmark snapshot or the dataset appears stale.
- Benchmark metric direction is opposite of your decision objective.
- The score difference between options is narrow and likely within implementation variance.
Current modeled score band for tracked entries is roughly 65.3 – 60.7.
Window
May 28
single timestamp
Mean delta
No trend
need another snapshot
Coverage
13
models in latest snapshot
Leaderboard preview (top 5)
- 1. Claude Opus 4.665.3
- 2. GLM-562.8
- 3. GLM-5.162.7
- 4. DeepSeek V3.260.9
- 5. Claude Sonnet 4.660.7