PRDBench - DEBUG: leaderboard
The second PRDBench round: after the judge agent's feedback on its first attempt, each model's minimal agent revises its 50 projects and the judge re-runs the test plans; the score is the average pass rate. It measures repair from feedback, where the headline board measures building from scratch.
Metric: Average pass rate (%). Source: github.com. Saturation forecast: Around February 2027. 8 models tracked. Updated 17 Nov 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 69 | #185 |
| 2 | Claude Sonnet 4.5 | 56.4 | #258 |
| 3 | Qwen 3 Coder 480B A35B Instruct | 48.29 | #653 |
| 4 | Kimi K2 0905 | 36.17 | #564 |
| 5 | GLM-4.7 | 35.33 | #379 |
| 6 | Gemini 3 Pro | 27.28 | #145 |
| 7 | DeepSeek V3.2 | 24.8 | #417 |
| 8 | MiniMax-M2 | 24.75 | #687 |
No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.
Interactive version: theaggregate.ai/benchmark?slug=prdbench-debug · How It Works · Data refreshed daily, snapshot 2026-11-17.