Google Antigravity 2.0 Wins Pantheon Build-Off in OpenSCAD LLM Benchmark

ModelRift pitted six AI coding agents — Codex 5.5 High, Claude Sonnet, Claude Opus, Cursor Composer, Google Antigravity 2.0, and ModelRift’s own system — against a single architectural task: generate an OpenSCAD script for the Roman Pantheon from two reference images, iterating via the OpenSCAD CLI to render PNG previews. The Pantheon was chosen deliberately as a mid-difficulty target that exercises OpenSCAD’s strengths in radial symmetry, Boolean operations, and parametric construction without straying into the organic geometry where the language struggles.

Google’s freshly released Antigravity 2.0 paired with Gemini 3.5 Flash High produced the strongest fully autonomous result. Where rivals eyeballed the references, Antigravity went looking for real Pantheon dimensions and fed them into the script as parametric values, even proposing a cutaway toggle to expose the interior coffers and niches. The migration from Antigravity 1.0’s VS Code-based IDE to a Codex-Desktop-style agentic app drew user backlash during launch week, but the model output itself impressed.

Client ergonomics shaped the runs as much as raw model quality. Codex Desktop’s habit of surfacing the loaded reference images and preview renders inline made visual CAD work easy to follow, while Claude Code’s terminal-first flow obscured the iteration loop. Cursor had the snappiest cycle but weaker geometry. Every agent handled the local OpenSCAD toolchain fine — the bottleneck was spatial judgment, not tool access, and even the winning model fell well short of a faithful Pantheon.