SWE Refactor Bench
Explore score history ↗20 whole-repository migrations only, so one task is 5 points of the mean score; the composite score gates on Migration Audit and all-or-nothing Behavioural Tests before Agentic Verification adds partial credit, and the harness and effort level are part of the result, so scores from different harnesses or effort levels are not the same measurement.
- Released
- 2026-08-24
- Evaluates
- coding agent
- Openness
- unknown
- Importer
- Model reports
SWE Refactor Bench paper
SWE Refactor Bench results and reported scores
Reported benchmark scores · Model reports
Reported scores: 26
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| Claude Opus 5 | Anthropic | 34.5 | 2026-08-24 | evidence |
| Claude Opus 5 | Anthropic | 20.5 | 2026-08-24 | evidence |
| Claude Opus 5 | Anthropic | 31.0 | 2026-08-24 | evidence |
| Claude Opus 5 | Anthropic | 28.5 | 2026-08-24 | evidence |
| Claude Opus 5 | Anthropic | 47.0 | 2026-08-24 | evidence |
| Claude Sonnet 5 | Anthropic | 6.0 | 2026-08-24 | evidence |
| Claude Sonnet 5 | Anthropic | 4.0 | 2026-08-24 | evidence |
| Claude Sonnet 5 | Anthropic | 8.5 | 2026-08-24 | evidence |
| Claude Sonnet 5 | Anthropic | 15.0 | 2026-08-24 | evidence |
| Claude Sonnet 5 | Anthropic | 9.0 | 2026-08-24 | evidence |
| DeepSeek V4 Flash | DeepSeek | 7.0 | 2026-08-24 | evidence |
| GLM 5.2 | Z.ai | 6.5 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 4.0 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 4.0 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 10.5 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 0.0 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 4.0 | 2026-08-24 | evidence |
| GPT-5.6 Luna | OpenAI | 5.5 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 19.0 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 7.0 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 28.5 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 6.5 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 4.0 | 2026-08-24 | evidence |
| GPT-5.6 Sol | OpenAI | 9.5 | 2026-08-24 | evidence |
| Kimi K3 | Moonshot AI | 19.5 | 2026-08-24 | evidence |
| Qwen 3.8 Max | Qwen | 10.0 | 2026-08-24 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.
Which sources cite SWE Refactor Bench?
- SWE Refactor Bench results table Einsia.AI, 2026-08-24