PRDBench: leaderboard
Each model drives the benchmark's minimal coding agent to build 50 Python projects from product requirement documents in one zero-shot round; a fine-tuned judge agent runs each project's test plan and the score is the average pass rate. Unlike unit-test benchmarks such as SWE-bench, the criteria mix unit tests with shell-interaction and file-comparison checks run by an agent judge.
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 | Claude Sonnet 4.5 | 69.19 | #258 |
| 2 | GPT-5.2 | 62.49 | #185 |
| 3 | Qwen 3 Coder 480B A35B Instruct | 43.84 | #653 |
| 4 | DeepSeek V3.2 | 40.11 | #417 |
| 5 | GLM-4.7 | 38.39 | #379 |
| 6 | Gemini 3 Pro | 22.76 | #145 |
| 7 | Kimi K2 0905 | 20.52 | #564 |
| 8 | MiniMax-M2 | 17.6 | #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 · How It Works · Data refreshed daily, snapshot 2026-11-17.