May 22, 2026
Dome sweet dome drama
Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark
It built a famous Roman temple, but commenters say: cool story, now fix the basics
TLDR: Antigravity 2.0 topped a test where AI tools had to build a 3D version of the Pantheon, showing how far this kind of design AI has come. But commenters turned it into a reality check, arguing that flashy wins mean less if the product is buggy, overhyped, or destined for the Google graveyard.
A new showdown asked several AI coding tools to recreate the Pantheon—yes, the giant ancient Roman landmark—in simple text-based 3D design code, and Antigravity 2.0 came out on top. On paper, that sounds like a big flex: this wasn’t just making a box or a tube, but building a recognizable domed structure with columns, steps, and that iconic front face. The company says this matters because better spatial reasoning means better 3D models people can actually use.
But the comments? Absolutely not ready to just clap politely. The loudest reaction was basically: Congrats on the trophy, but can the product work first? One frustrated user dragged Antigravity for forcing a browser login every time and for an app that “won’t update at all,” which instantly turned the victory lap into a mini roast. Another commenter summed up the whole AI mood in 2026: the tech is improving so fast that the “goalposts” are sprinting away. In other words, people know this is impressive—they’re just no longer impressed for more than five seconds.
Then came the classic big-tech trust issues. One hot take predicted Antigravity could be beaten next month and sunset by Google next year, because if it doesn’t mint billions, why get attached? Others tossed cold water on the idea that this means human design software is doomed anytime soon, pointing out that even major design tools’ own AI assistants are still pretty rough. And one of the more interesting side debates asked: why aren’t we seeing more specialized AI builders for design, instead of one giant model trying to do everything from code to poetry to architecture? So yes, Antigravity won the benchmark—but the real event was the crowd yelling, joking, and side-eyeing the whole industry from the cheap seats.
Key Points
- •ModelRift ran a benchmark that gave several AI coding tools the same architectural task: generate an OpenSCAD model of the Pantheon from reference images.
- •The benchmark was designed to test how well systems convert architectural reference material into parametric CAD code, using the OpenSCAD CLI for preview rendering and iteration.
- •The article argues that the Pantheon is a useful benchmark because it combines radial symmetry, repeated columns, a dome, an oculus, and a portico without requiring organic sculpting.
- •ModelRift presents OpenSCAD as a strong fit for LLM-generated geometry because it is plain-text code that supports inspectable, reproducible, and editable parametric structures.
- •The article contrasts OpenSCAD with UI-driven or tool-control workflows such as Blender MCPs, arguing that code-first geometry is more direct for CAD-like tasks.