Google Antigravity 2.0 Tops OpenSCAD Architectural LLM Benchmark

The Benchmark

ModelRift tested six AI coding tools on generating an OpenSCAD parametric 3D model of the Pantheon from two reference images and a minimal prompt, targeting genuine spatial reasoning in a complex geometric task.

Results

• Google Antigravity 2.0 — 4.5/5 autonomous quality (best). Researched actual Pantheon dimensions; the only model to include the interior coffered ceiling with five rings of 28 coffers
• ModelRift/Gemini Flash 3.0 (human-in-loop) — 3.8/5 overall best, 2× Antigravity's time
• Claude Sonnet — 3.4/5. Cleanest autonomous proportions
• Codex 5.5 High — 3.0/5, with a preview-vs-mesh mismatch
• Claude Opus — 3.0/5
• Cursor Composer — fastest, worst quality at 1.4/5

What Made Antigravity 2.0 Stand Out

Launched at Google I/O 2026, Antigravity 2.0 is a Gemini 3.5 Flash High-powered agent-first desktop app. Where other tools modeled only the facade, Antigravity independently researched real Pantheon dimensions and created a cutaway mode exposing both interior and exterior detail.

Key Takeaways

Speed did not correlate with quality. Visual feedback loops outperformed pure text iteration. Preview correctness is not mesh correctness — a subtlety that caught Codex off guard and matters greatly for production 3D workflows.