RepoReasoner: leaderboard
Output Prediction Pass@1 on original repository data with a 10k-token Oracle context. Scores are fractions from 0 to 1; retrieval contexts and I/O-rewritten data are separate protocols.
Metric: Pass@1 (0-1, Output Prediction, original data, 10k Oracle). Source: arxiv.org. Saturation forecast: Around January 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek R1 | 0.69 | #251 |
| 2 | GPT-4.1 Mini | 0.67 | #355 |
| 3 | Gemini 2.5 Flash | 0.62 | #245 |
| 4 | Qwen 3 235B A22B | 0.61 | #313 |
| 5 | Qwen 2.5 Coder 14B | 0.52 | #1184 |
| 6 | Qwen 2.5 14B | 0.51 | #725 |
| 7 | DeepSeek R1 Distill Qwen 14B | 0.47 | #844 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=reporeasoner · How It Works · Data refreshed daily, snapshot 2026-10-20.