In a spatial reasoning benchmark published by 3D platform ModelRift, Google Antigravity 2.0 powered by Gemini 3.5 Flash High led all autonomous models with a 4.5 out of 5 quality score. The benchmark tasked AI tools with converting architectural reference images into parametric OpenSCAD code to reconstruct Rome's Pantheon. The findings reveal that text-based parametric code generation offers a clearer pathway for AI in spatial engineering than complex UI automation.
Benchmarking Architectural CAD Generation
ModelRift evaluated how well leading AI coding systems convert reference imagery into functional 3D models. OpenSCAD was chosen as the target language because its compact, text-first syntax uses constructive solid geometry and parameters. Language models can describe complex spatial objects as nested transformations, loops, and Boolean operations far more effectively than driving a 3D application UI.
The benchmark required AI agents to recreate the Pantheon—a structure featuring a radial rotunda, dome, central oculus, rectangular portico, and pediment. Agents had access to a local OpenSCAD CLI to render PNG previews, allowing them to inspect and iterate on their code before outputting final files.
Benchmark Results and Model Performance
Performance across six test configurations varied significantly in terms of visual quality and execution speed:
- Google Antigravity 2.0 / Gemini 3.5 Flash High: 4.5/5 quality, 1/5 speed (~12 min). The best autonomous result, utilizing exact dimensions and complex interior detailing.
- ModelRift / Gemini Flash 3.0: 3.8/5 quality, 1/5 speed (~10 min). The top human-in-the-loop score, driven by visual annotation feedback on renders.
- Claude Code 2.1 / Sonnet 4.6: 3.4/5 quality, 1/5 speed. Delivered clean massing and balanced overall proportions.
- Codex 5.5 High: 3.0/5 quality, 4/5 speed. Produced high detail including entablature inscriptions, but was hampered by an export mismatch between the PNG preview and the final STL file.
- Claude Code 2.1 / Opus 4.7: 3.0/5 quality, 2/5 speed. Showed clear portico structure but lacked color discipline.
- Cursor 3.5 / Composer 2.5: 1.4/5 quality, 5/5 speed. Delivered the fastest interaction loop but generated the lowest visual quality.
Google Antigravity 2.0 was the only autonomous system to search for authentic architectural parameters, rendering an accurate 5-ring, 28-coffer ceiling inside the dome. It also introduced a cutaway toggle in the code to reveal interior details.
Trade-offs: Latency, Cost, and User Workflows
While Antigravity 2.0 achieved impressive spatial quality, the system comes with distinct trade-offs. The test run took approximately 12 minutes to complete, making it one of the slowest autonomous agents in the evaluation.