BenchCAD through AgentModel
BenchCAD is a benchmark for programmatic CAD. Its Vision2Code task gives a model four rendered views of a mechanical part — a 2×2 composite, nothing else — and asks it to write a program that rebuilds the part. The dataset ships the ground-truth code for every part, so an answer can be scored on geometry rather than on how the code reads.
These are BenchCAD's images and BenchCAD's prompt, run through the real AgentModel app: the same chat path a person uses, OpenSCAD out the other end, rendered exactly as it came. Every model is then scored with BenchCAD's own metric — voxel IoU against their ground truth, failures counted as zero. 36 parts, one per family, 12 per difficulty tier. That is a small slice of their 17,900, so these numbers say how AgentModel does on this task; they are not comparable to the published leaderboard.
Part images and ground-truth code from the BenchCAD dataset, used under CC-BY-4.0. Harness: BenchCAD-main. Paper: arXiv:2605.10865.