RepairBench: leaderboard
LLM-driven program repair benchmark: 46 frontier models tested on 835 real bugs from Defects4J and GitBug-Java, measuring plausible patch generation at pass@1.
Metric: Plausible@1 (%). Source: repairbench.github.io. Saturation forecast: Estimated already saturated. 39 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | O4 Mini (2025-04-16) (High) | 50.3 | #179 (O4 Mini (2025-04-16)) |
| 2 | O3 Mini (2025-01-31) (High) | 46.4 | #240 (O3 Mini (2025-01-31)) |
| 3 | DeepSeek R1 | 45.2 | #251 |
| 4 | Claude 3.7 Sonnet (20250219) | 44 | #201 |
| 5 | Claude 3.5 Sonnet (20241022) | 41.8 | #291 |
| 6 | GPT-4.1 | 41.3 | #248 |
| 7 | Gemini 2.5 Flash (Preview 05-20) | 40.6 | #216 |
| 8 | DeepSeek V3 (0324) | 39.6 | #340 |
| 9 | Claude 3.5 Sonnet (20240620) | 39.1 | #299 |
| 10 | Gemini 2.5 Pro (Preview 03-25) | 38.3 | #123 |
| 11 | DeepSeek V3 | 37.1 | #321 |
| 12 | Gemini 1.5 Pro (002) | 33.2 | #349 |
| 13 | GPT-4o (2024-11-20) | 32.6 | #381 |
| 14 | Mistral Medium 3 | 32.1 | #500 |
| 15 | GPT-4o (2024-08-06) | 31.7 | #336 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=repairbench · How It Works · Data refreshed daily, snapshot 2026-10-20.