What happens when language models write CAD? Each benchmark gives the same requests (written specs, dimensioned drawings, reference photos) to several models and tools. Every result is rendered the same way and scored by the same judge, and the model files are published so you can check them yourself.
Opus 5.5 as a CAD agent vs one-shot Gemini 3.7/3.8 Flash and GPT-6 Astra on six parts from drawings and short specs, with judge scores, time and cost.
Opus 5.5 as a CAD agent vs one-shot Gemini 3.7 and 3.8 Flash on eight harder parts: enclosures, a hinge, gears, a C-clamp and models from photos.
Opus 5.5 builds six sheet-metal parts with flat patterns, once in OpenSCAD and once in SolidCraft. Folded parts, flat patterns, judge scores and DXFs.
AgentModel