Sphinx - Java: leaderboard
Sphinx checklist-coverage score on its 500 Java pull requests, 450 needing changes and 50 bug-free controls whose correct review is no comment: a GPT-4o judge checks each generated review against a checklist derived from the merged change, buggy cases weighted 0.9; percent, higher is better. Microsoft and University of Rochester, 2026.
Metric: Checklist Coverage Score (%; GPT-4o judge, 450 buggy + 50 bug-free PRs per language). Source: arxiv.org. Saturation forecast: Estimated already saturated. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 | 26.46 | #492 |
| 2 | Gemini 2.5 Pro | 26.31 | #262 |
| 3 | Claude 3.7 Sonnet | 25.92 | #455 |
| 4 | Qwen 2.5 72B Instruct | 24.94 | #1093 |
| 5 | Phi-4 | 24.53 | #1920 |
| 6 | O3 Mini | 23.84 | #536 |
| 7 | O4 Mini | 23.82 | #354 |
| 8 | Qwen 3 235B A22B | 23.36 | #622 |
| 9 | Claude 3.5 Sonnet | 23.07 | #747 |
| 10 | GPT-4o (2024-08-06) | 22.93 | #731 |
| 11 | GPT-4o Mini | 22.56 | #1627 |
| 12 | DeepSeek V3 | 22.34 | #705 |
| 13 | Qwen 3 30B A3B | 21.96 | #1124 |
| 14 | GPT-4.1 Nano | 21.88 | #1992 |
| 15 | GPT-4.1 Mini | 21.57 | #775 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sphinx-java · How It Works · Data refreshed daily, snapshot 2026-11-12.