CurveBench - Easy: leaderboard
Metric: Tree-Generation Accuracy (%; exact containment tree, 45 held-out test images, four rollouts each). Source: arxiv.org. Saturation forecast: Around December 2026. 8 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 71.1 |
| 2 | Claude Opus 4.5 | 32.2 |
| 3 | GPT-5.4 | 30.6 |
| 4 | GPT-5.4 Mini | 13.9 |
| 5 | Qwen 3 VL 8B (Thinking) | 2.8 |
| 6 | Qwen 3 VL 8B Instruct | 1.7 |
Interactive version: theaggregate.ai/benchmark?slug=curvebench-easy · How It Works · Data refreshed daily, snapshot 2026-09-25.