Antigravity 2.0 Tops OpenSCAD Architectural 3D LLM Benchmark

ModelRift ran a benchmark testing multiple AI coding tools on implementing the Pantheon temple in OpenSCAD. Antigravity 2.0 (Gemini 3.5 Flash High), announced at Google I/O 2026, took first place in the autonomous execution category with a score of 4.5/5.

Benchmark Design

The task was to generate an OpenSCAD file including the Pantheon's rotunda, dome, portico, columns, and pediment from 2 reference images, rendering PNG previews via the OpenSCAD CLI and iterating for improvement. The Pantheon combines radial symmetry, Boolean operations, and extruded shapes, making it well-suited to OpenSCAD while also requiring architectural understanding.

Key Results

• Antigravity 2.0 + Gemini 3.5 Flash High: 4.5/5, about 12 minutes. It searched for and used actual Pantheon dimensions, included the inscription, and was the only one to implement the interior coffered ceiling pattern. It mathematically subtracted a structure of 5 rings × 28 coffers.

• Codex 5.5 High: 3.0/5. Detail density such as the inscription was high, but the score dropped due to inconsistency between the PNG preview and the final STL export.

• Claude Code + Sonnet 4.6: 3.4/5, the slowest. Clean massing and balanced proportions.

• Claude Code + Opus 4.7: 3.0/5. A clear portico and stepped base.

• Cursor 3.5 / Composer 2.5: 1.4/5, the fastest. The weakest output, lacking proportion and detail.

• ModelRift + Gemini Flash 3.0 (with human intervention): 3.8/5. Used an iterative annotation workflow.

Why OpenSCAD Was Chosen

OpenSCAD expresses geometric structures as plain-text code, letting LLMs directly write commands like "repeat 28 columns around a radius." Compared to tool-control approaches like Blender MCP, it involves less indirection for CAD work, and the resulting output can be inspected, reproduced, and modified.

Workflow Differences

Codex Desktop displayed reference images, code edits, and previews in the same thread within the conversation, making it easy to track visual context. Cursor offered the fastest iteration loop, but its output quality lagged behind. Claude Code was terminal-centric, with visual elements being less explicit.