Antigravity 2.0 Tops OpenSCAD Architectural 3D Benchmark – ModelRift Tests 6 LLMs on the Pantheon

ModelRift ran a practical benchmark: they asked six AI coding tools to build the Pantheon in OpenSCAD from reference images. The goal was to test how well each system turns architectural reference material into parametric CAD code. The prompt used two images (front facade and aerial view) and required using the OpenSCAD CLI to preview and iterate.

Why Pantheon + OpenSCAD?

Basic prompts like "cube with a hole" test only simple syntax (difference, cube, cylinder). The Pantheon sits in a middle ground: it has radial symmetry (rotunda, dome, oculus), straight portico faces, columns, stepped bases, and a triangular pediment. This mix tests an LLM's ability to handle nested transformations, Boolean operations, loops, and named modules — all native to OpenSCAD's plain-text representation. OpenSCAD keeps geometry as the artifact, avoiding the indirection of Blender MCPs or UI actions.

Benchmark Results

Six systems were tested. Each output was scored on quality (1-5) and given a summary. The table below shows the top results:

Tool & Model	Time	Quality	Summary
Antigravity 2.0	●●●○○ (3/5)	●●●●○ (4.5/5)	Best quality. Captured Pantheon proportions, dome with oculus, portico, columns, pediment, and facade details. Architecture most faithful to references.
Codex 5.5 High	●●●●○ (4/5)	●●●○○ (3.0/5)	Strong detail density, including inscription on entablature. But final STL didn't match PNG preview, holding score down.
Cursor 3.5 / Composer 2.5	●●●●● (5/5)	●○○○○ (1.4/5)	Fastest run but weakest output: poor proportions, color discipline, and architectural details.

Full results include three more entries (not detailed here). The benchmark code and render comparisons are available on the original post.

Practical Takeaways

• Antigravity 2.0 produced the most architecturally accurate OpenSCAD code, with correct dome rings, column spacing, and facade relationships.
• Codex 5.5 added fine details (inscription) but suffered an export mismatch — the preview looked better than the final STL.
• Cursor 3.5 was quick but the geometry was crude; it's fine for rapid prototyping but not for production CAD.
• The benchmark confirms OpenSCAD is a strong target for LLM-generated geometry: plain text, compact vocabulary, and easy iteration via CLI.

If you're using AI coding agents for parametric 3D modeling, especially for architectural or mechanical parts with radial symmetry and Boolean operations, this benchmark gives a clear signal: Antigravity 2.0 currently leads in quality. For speed-first tasks, Cursor 3.5 might still be useful if you're willing to iterate heavily.