PBT-Bench (Baseline Prompt): leaderboard
PBT-Bench (2026) gives an OpenHands coding agent 100 Python library problems with injected semantic bugs and asks it to write tests that expose them; this board uses the open-ended baseline prompt without property-based testing scaffolding. Bug recall in percent, mean of three runs; higher is better.
Metric: Bug recall (%) on the 100 problems under the open-ended baseline prompt (write tests that expose the injected bugs, no property-based testing scaffolding), OpenHands agent scaffold, mean of three runs, each problem weighted equally; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked. Updated 7 Oct 2026.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Sonnet 4.6 | 76.7 | #173 |
| 2 | GLM-5.1 | 66.6 | #198 |
| 3 | DeepSeek V3.2 | 64.2 | #419 |
| 4 | Gemini 3 Flash | 56.2 | #175 |
| 5 | Qwen 3.6 Plus | 53.5 | #238 |
| 6 | Grok 4.1 Fast | 50.1 | #427 |
| 7 | Step 3.5 Flash | 37.8 | #526 |
| 8 | PBT-Bench Qwen 3.5-30B-A3B (checkpoint unspecified) | 31.4 |
No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.
Interactive version: theaggregate.ai/benchmark?slug=pbt-bench-baseline-prompt · How It Works · Data refreshed daily, snapshot 2026-11-21.