A practical benchmark comparing six AI coding tools (Codex 5.5 High, Claude Sonnet, Claude Opus, Cursor Composer, Google Antigravity 2.0, and ModelRift) on a single task: generating a 3D model of the Pantheon in OpenSCAD from reference images. Google Antigravity 2.0 with Gemini 3.5 Flash High produced the best fully autonomous result (4.5/5), notably implementing the dome's interior coffered ceiling pattern and using real architectural measurements. ModelRift with human-in-the-loop annotation scored 3.8/5. Cursor was fastest but weakest (1.4/5). Codex showed strong spatial reasoning but suffered from a preview-to-STL export mismatch. Key takeaways: speed doesn't predict quality, tool access wasn't the bottleneck, and human visual feedback still meaningfully improves results over fully autonomous generation for complex spatial CAD tasks.
