Benchmark Compares CadQuery and OpenSCAD for AI-Driven CAD Work

Benchmark Overview

ModelRift performed a controlled comparison between CadQuery and OpenSCAD to evaluate their suitability for AI-driven CAD work. The benchmark focused on which tool an AI agent could use to generate a correct, printable, and functional part without human intervention. The study used six Claude Opus 5 agents to complete three distinct CAD tasks with each tool.

Methodology and Setup

Each agent operated independently, with tool versions CadQuery 2.8.0 (Python 3.14) and OpenSCAD 2026.06.12 running on an M-series Mac. ModelRift's internal `openscad-skill` was used for OpenSCAD, and a ported version of this skill was created for CadQuery, including an offscreen renderer and B-rep validity metrics. Both setups received identical 3D-printing design rules, wall thickness, clearances, and overhang advice. Agents were capped at 12 versions and required to report all failures with verbatim error text.

Results and Observations

All six generated parts were printable, indicating that both CadQuery and OpenSCAD possess the core capability for AI-driven part creation. The primary distinction between the tools emerged in their failure characteristics rather than their ability to complete tasks. The study did not detail the specific failure modes but noted this as the key differentiator.

Tool Specifics

The CadQuery setup included an API cheatsheet, which was not necessary for OpenSCAD as the model already had strong familiarity with its syntax. This provided CadQuery an advantage in syntax handling but did not affect geometric outcomes. The resulting STL file sizes were comparable: 12.4 KB for OpenSCAD and 13.1 KB for CadQuery.