3DCodeBench - Executability: leaderboard
Procedural 3D modeling by code (3DCodeBench, Gao et al., 2026): for 212 Infinigen object categories the model writes a standalone Blender 5.0 Python script from a text description or from four reference renders, in one call. The score is the percentage of scripts that run and produce a mesh, averaged over the text-to-3D and image-to-3D tracks, with each model at its best thinking level; higher is better.
Metric: Executability (%; Blender 5.0 scripts that run and yield a mesh, 212 instances, mean of the text-to-3D and image-to-3D tracks, single call at the best thinking level). Source: arxiv.org. Saturation forecast: Around February 2027. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4.7 | 91 | #75 |
| 2 | GPT-5.5 (High) | 90.6 | #48 (GPT-5.5) |
| 3 | GPT-5.4 (Medium) | 86.6 | #121 (GPT-5.4) |
| 4 | Claude Sonnet 4.6 | 80.4 | #158 |
| 5 | GPT-5.4 Mini (Medium) | 73.1 | #389 (GPT-5.4 Mini) |
| 6 | Gemini 3.1 Pro (Preview) | 72.5 | #91 |
| 7 | Gemma 4 31B | 58.2 | #393 |
| 8 | Gemini 3.1 Flash Lite | 57.5 | #363 |
| 9 | Gemma 4 26B | 51.7 | #614 |
| 10 | Claude Haiku 4.5 | 50.2 | #568 |
| 11 | Gemini 3 Flash | 48.1 | #159 |
| 12 | Gemini 3.5 Flash | 46.4 | #98 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=3dcodebench-executability · How It Works · Data refreshed daily, snapshot 2026-11-12.