CAD Arena
Agents
Normal Research's CAD Arena asks an agent to rebuild a part from a manufacturer-licensed engineering drawing as editable CAD through a platform scripting API, with 18 drawings on five platforms (NX Open, SolidWorks, Onshape FeatureScript, Autodesk Fusion and Build123d) for 90 cases. Each case scores S = G x (0.7 + 0.3 E), where G is the FrontierCAD verifier's 0-1 geometric match of the final STEP solid to the reference part and E is a judge's 0-1 editability score for the native sketches, features and parameters; the score is the mean over scored cases on a 0-100 scale, not a pass rate. Models run in their provider's own agent harness where one exists, so results measure model, harness and environment together.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 74.6 |
| GPT-6 Astra | 67.1 |
| Claude Fable 5.1 | 66.2 |
| GPT-6 Sol | 52.4 |
| Gemini 3.8 Flash | 41.0 |
| Grok 4.6 | 34.8 |
| Muse Spark 1.3 | 31.8 |
| GPT-6 Luna | 29.8 |
| Grok 4.7 | 25.8 |
| DeepSeek V4 Flash Vision Exp | 18.7 |
| Inkling-Small | 14.3 |
| Inkling | 12.4 |