BenchCAD 1.0 - Vision2Code: leaderboard
The model sees four orthographic views of a mechanical part and writes a CadQuery program; the program is executed and its solid compared with the ground truth by voxel IoU, with parts that fail to run scoring 0. These are the BenchCAD maintainers' own graded runs on the 1.0 dataset, not the vendor-reported figures some launch posts quote, and not BenchCAD 2.1 Core's six-task agent cases.
Metric: IoU-score (voxel IoU x execution rate, 0 to 1), image to CadQuery code, re-graded by BenchCAD. Source: benchcad.com. Saturation forecast: Around May 2028. 8 models tracked. Updated 16 Nov 2026.
Top models
| # | Model | Setting | Score | Overall rank |
|---|---|---|---|---|
| 1 | Kimi K3 | Max | 0.367 | #56 |
| 2 | Grok 4.6 | xHigh | 0.3638 | #45 |
| 3 | Grok 4.5 | High | 0.3194 | #85 |
| 4 | Gemini 3.1 Pro (Preview) | Thinking | 0.289 | #93 |
| 5 | Claude Opus 4.7 | Max | 0.2692 | #78 |
| 6 | Claude Sonnet 4.6 | Max | 0.222 | #165 |
| 7 | GPT-4o | 0.1823 | #744 | |
| 8 | O3 | 0.1218 | #228 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=benchcad-1-0-vision2code · How It Works · Data refreshed daily, snapshot 2026-11-16.