CATTest - Bug Discovery P-score: leaderboard
Fine-grained P-score for how trustworthy a VLM testing agent's bug reports are when matched against annotated ground-truth bugs across 102 AI-generated web applications; per-project scores are summed and normalized (CATTest, arXiv 2609.00081v1, Table 1). 0-100; higher is better.
Metric: P-score (0-100; precision-oriented bug reporting). Source: arxiv.org. Saturation forecast: Around April 2027. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 48.56 | #64 |
| 2 | Claude Opus 4.6 | 48.2 | #67 |
| 3 | GPT-5.5 | 46.88 | #33 |
| 4 | Claude Sonnet 4.6 | 41.08 | #94 |
| 5 | Claude Opus 4.7 | 38.3 | #51 |
| 6 | Qwen 3.6 27B | 37.92 | #182 |
| 7 | GLM-5V Turbo | 33.5 | #172 |
| 8 | Qwen 3.5 Plus | 32.62 | #133 |
| 9 | Kimi K2.6 | 29.43 | #113 |
| 10 | Qwen 3.5 27B | 26.72 | #192 |
| 11 | Seed 2.0 Pro | 26.55 | #122 |
| 12 | GPT-5.4 Mini | 25 | #228 |
| 13 | Gemini 3.1 Flash Lite | 23.37 | #200 |
| 14 | Seed 2.0 Lite | 23.2 | #161 |
| 15 | Gemini 3 Flash | 22.99 | #103 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cattest-bug-discovery-p-score · How It Works · Data refreshed daily, snapshot 2026-11-05.