PAIR-Bench: leaderboard
PAIR-Bench (University of Texas at Dallas, University of Central Florida and FPT Software, 2026) evaluates code improvement on 440 wrong-answer Codeforces Python programs, with progressive hints that target grouped failure scenarios at increasing depth. The final fix rate is the percent of programs that pass every hidden test within a 10-turn hint budget; higher is better.
Metric: Final fix rate (%; share of the 440 wrong-answer Codeforces Python programs repaired to pass every hidden test within a 10-turn budget of progressive hints generated by GPT-OSS 120B, at most 3 turns per failure scenario; higher is better). Source: arxiv.org. Saturation forecast: Estimated already saturated. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 | 99.31 | #406 |
| 2 | Gemini 2.5 Flash Lite | 95.45 | #958 |
| 3 | Qwen 3 Coder 30B A3B Instruct | 90.9 | #972 |
| 4 | Llama 3.3 70B Instruct | 85.9 | #1357 |
| 5 | Ministral 3 14B | 85.9 | #1845 |
| 6 | GPT-4o Mini | 85.45 | #1623 |
| 7 | Mistral Small 3.2 | 79.77 | #1591 |
| 8 | Gemma 3 27B | 72.04 | #1631 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pair-bench · How It Works · Data refreshed daily, snapshot 2026-11-13.