Django Benchmark Suite - Program Repair: leaderboard
Part of a Django benchmark suite (Shibaev et al., 2026): the model fixes a broken method given its file and the failing Django test stack trace. Pass@1 in percent over 359 held-out methods, counted when all tests mapped to the method pass; higher is better.
Metric: pass@1 (%; Program repair: the model fixes a broken method from the file and the Django test stack trace; the output is inserted into the file and counts only when every Django test mapped to the method passes; 359 held-out methods of a leakage-controlled Django repository snapshot; greedy decoding, no agent scaffold). Source: arxiv.org. Saturation forecast: Around February 2027. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 79.67 | #167 |
| 2 | Claude Opus 4.5 | 78.27 | #156 |
| 3 | Claude Sonnet 4.5 | 69.36 | #250 |
| 4 | Gemini 2.5 Pro | 67.97 | #266 |
| 5 | Gemini 3 Flash | 62.95 | #164 |
| 6 | GPT-4.1 | 61.84 | #495 |
| 7 | Claude Haiku 4.5 | 61.56 | #568 |
| 8 | Gemini 3 Pro | 61 | #137 |
| 9 | Qwen 2.5 Coder 32B Instruct | 47.35 | #1562 |
| 10 | Qwen 3 32B | 46.8 | #960 |
| 11 | Qwen 2.5 Coder 7B Instruct | 17.83 | #3852 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=django-benchmark-suite-program-repair · How It Works · Data refreshed daily, snapshot 2026-11-13.