Sphinx: leaderboard
Sphinx (Microsoft and University of Rochester, 2026) asks models to write review comments for 2,500 merged single-file GitHub pull requests in five languages, given the diff, PR description, linked issues and code; a tenth are bug-free controls whose correct review is no comment. A GPT-4o judge checks the review against a checklist derived from the differences between a model-written pseudo solution and the merged code, weighting buggy cases 0.9 and bug-free cases 0.1; percent, higher is better. The Sphinx models are the authors' fine-tunes on the Sphinx training set.
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 | 34.23 | #492 |
| 2 | Gemini 2.5 Pro | 32.75 | #262 |
| 3 | O4 Mini | 32.01 | #354 |
| 4 | Claude 3.7 Sonnet | 31.39 | #455 |
| 5 | O3 Mini | 30.84 | #536 |
| 6 | Qwen 2.5 72B Instruct | 30.53 | #1093 |
| 7 | Claude 3.5 Sonnet | 29.81 | #747 |
| 8 | Qwen 3 235B A22B | 29.67 | #622 |
| 9 | GPT-4o (2024-08-06) | 29.4 | #731 |
| 10 | Qwen 3 30B A3B | 29.14 | #1124 |
| 11 | GPT-4o Mini | 28.84 | #1627 |
| 12 | GPT-4.1 Mini | 26.74 | #775 |
| 13 | DeepSeek V3 | 26.45 | #705 |
| 14 | Qwen 3 32B | 26.06 | #960 |
| 15 | DeepSeek R1 | 25.6 | #467 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=sphinx · How It Works · Data refreshed daily, snapshot 2026-11-12.