AI Agent CAD Benchmark: CadQuery's Explicit Failure vs OpenSCAD's Silent Failure

ModelRift conducted a benchmark comparing the unmanned CAD part generation performance of CadQuery and OpenSCAD using Claude Opus 5 and Claude Code. While the two tools showed no difference in their ability to produce final printable (watertight) parts, they displayed a clear contrast in failure modes and verifiability.

CadQuery fails explicitly and early by raising exceptions, but its validity checks during boolean operations may be incomplete. In contrast, OpenSCAD tends to exhibit silent failures by certifying incorrect geometry as 'clean' without errors or warnings, making it more critical in unmanned environments. In fact, there was a case where OpenSCAD's Manifold backend returned NoError for an STL file containing non-manifold edges. Since self-reporting is unreliable for both tools, independent verification via direct mesh parsing is essential.

In terms of performance, OpenSCAD's geometry recompute speed is 30–100 times faster than CadQuery's (16 ms vs 1.9 s), but this difference is negligible compared to model inference time. OpenSCAD was more efficient in both total execution time and token usage (OpenSCAD 2061s/297k tokens vs CadQuery 2406s/347k tokens), and its code was also more concise. CadQuery, being B-rep based, facilitates parameter validation, whereas OpenSCAD, resulting in a triangle soup form after CSG, makes outline recognition difficult.

Ultimately, ModelRift decided to retain OpenSCAD considering structural advantages such as WASM sandboxing, rebuild speed, and multi-color 3MF export. However, to compensate for the limitations of visual feedback, a numeric-based verification (Numeric Check) process was mandatorily introduced to check wall thickness, gaps, interference volume, etc.