Hacker News
Six Claude Opus 5 Agents Benchmarked on CAD: OpenSCAD Produced Five Silent Geometry Failures, CadQuery Broke Loudly Instead
ModelRift's September 12 benchmark put six Claude Opus 5 agents through Claude Code on three 3D-printing tasks (a gusseted L-bracket, a two-part snap-fit PCB enclosure with M2 posts and a USB-C cutout, and an M24x2 threaded hose-barb adapter), each built twice, once per tool. Both tools took 11 versions; OpenSCAD used 473 lines to CadQuery's 573, reported 2 errors to CadQuery's 5, and produced 5 silent geometry failures including corrupted exports during T2 that no self-report caught. The takeaway generalizes past CAD: CadQuery recomputes in 1.93s versus OpenSCAD's 16ms, and the slower tool that fails loudly beat the fast one that fails silently, because the agent cannot debug what the tool never reports.
↳ Follow the thread