RepoProbe: leaderboard
Overall Performance percentage on 500 repository-understanding questions across 50 repositories using a fixed Claude Code harness. This is the paper’s combined overall metric, distinct from Knowledge Score and Perfect Solve Rate.
Metric: Overall Performance (%). Source: arxiv.org. Saturation forecast: Around January 2027. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 62.7 | #110 |
| 2 | Claude Opus 4.6 | 62.1 | #63 |
| 3 | GPT-5.4 | 60.4 | #79 |
| 4 | Claude Sonnet 4.6 | 60.2 | #90 |
| 5 | Claude Opus 4.5 | 58.1 | #83 |
| 6 | Grok 4.20 | 56.8 | #175 |
| 7 | Gemini 3 Flash | 55.4 | #99 |
| 8 | GLM-5 | 54.8 | #141 |
| 9 | Gemini 3.1 Pro (Preview) | 54.1 | #58 |
| 10 | Kimi K2.5 | 53.7 | #144 |
| 11 | DeepSeek V3.2 | 52.9 | #205 |
| 12 | Gemini 3 Pro | 51.3 | #80 |
| 13 | GLM-4.7 | 50.7 | #192 |
| 14 | Qwen 3.5 Plus | 50 | #125 |
| 15 | Kimi K2 | 45.9 | #243 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=repoprobe · How It Works · Data refreshed daily, snapshot 2026-10-20.