CATTest - Pass/Not-Pass F1: leaderboard
Coarse-grained pass/not-pass F1 when a VLM testing agent judges whether each of 102 AI-generated web applications passes, under the CATJudge framework with a 250-turn budget at provider defaults (CATTest, arXiv 2609.00081v1, Table 1). Percent; higher is better.
Metric: F1 (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 17 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 89.7 | #51 |
| 2 | Claude Sonnet 4.6 | 89.57 | #94 |
| 3 | Claude Opus 4.6 | 88.2 | #67 |
| 4 | GPT-5.4 | 86.05 | #83 |
| 5 | GPT-5.5 | 85 | #33 |
| 6 | Gemini 3 Flash | 83.02 | #103 |
| 7 | Gemini 3.1 Pro (Preview) | 82.8 | #64 |
| 8 | Qwen 3.5 27B | 80 | #192 |
| 9 | Qwen 3.6 27B | 79.72 | #182 |
| 10 | Kimi K2.6 | 79.47 | #113 |
| 11 | Qwen 3.5 Plus | 78.08 | #133 |
| 12 | GPT-5.4 Mini | 70 | #228 |
| 13 | GLM-5V Turbo | 69.7 | #172 |
| 14 | Seed 2.0 Pro | 58.33 | #122 |
| 15 | Gemini 3.1 Flash Lite | 54.7 | #200 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=cattest-pass-not-pass-f1 · How It Works · Data refreshed daily, snapshot 2026-11-05.