LLMs Build 3D Scenes and Videos as Code: AA Tests GPT-6.2 Astra, Experts Urge Caution

mark_k · x · 2026-11-20

Artificial Analysis asks whether language models can make better videos than video generation models, testing GPT-6.2 Astra — OpenAI's current #1 on the AA-3D benchmark — by building the Sydney Opera House, an exploded mechanical watch, and a dragon attacking a medieval town as three.js code, alongside motion graphics, code-based video editing, and driving Blender.

markk pushes back: while impressive, results depend heavily on content type — photorealistic humans would be much harder to achieve via 3D rendering, suggesting real limits to this "LLMs as code renderers" approach.

Related event: Artificial Analysis Tests LLM-Generated Video Against Video Models(2 posts)→

Original post →

More from Models

Models channel →