How AA-3D Works: Models as Agents in Sandboxes Building 3D Scenes
ArtificialAnlys · x · 2026-11-10
Artificial Analysis explains how its AA-3D benchmark works (build, check, vote, rank):
- Each model works as an agent in an offline sandbox, turning a written brief into a three.js web page, with up to 3 hours and 250 turns
- The sandbox provides Blender, Rapier physics, Python, a headless browser for self-testing, plus a broad library of materials, 3D models and characters
- Human panelists compare blind pairs of scenes from the same brief, spending at least 30 seconds per scene before voting
- Votes are converted into Elo scores with 95% confidence intervals
Example brief: an animated Starship mission from launch to orbit, including booster tower catch and ship landing.
Related event: Claude Fable 5.2 Tops AA-3D Leaderboard with 67-Point Elo Lead(6 posts)→
More from Models
- Anthropic's paternalism and Claude's paternalism are the same phenomenon — amplifiedamp · 2026-11-10
- SemiAnalysis Launches GDPx Leaderboard: Claude Fable 5.2 Debuts on Top at 16.4% — scaling01 · 2026-11-10
- Anthropic caps GPU kernel optimization in Fable 5.2 training as a safeguard — community asks: pre or post nerf? — WolframRvnwlf · 2026-11-10
- Fable IS Mythos with more guardrails, users say as Anthropic calls them two models — repligate · 2026-11-10
- Researchers Test New Claude Models as Long-Horizon Research Assistants — chaitjo · 2026-11-10
- Open-source Mirage model trends on Hugging Face with zero-shot detection and VQA — Cactus-Compute · 2026-11-10