Its v1.0 evaluation covers 21 models and 55 briefs. Models operate as agents in an offline sandbox with Blender, Rapier physics, Python, a headless browser and a CC0 asset library; each run ends after three hours or 250 turns and produces an index.html scene. Recruited human raters compare pairs of scenes for the same brief without seeing which model made them, and the results are ranked by Elo.
Claude Sonnet 5.5 leads with 1194 Elo, but Artificial Analysis says the top four are statistically tied because their 95% confidence intervals overlap. The article also reports a sampled run-cost comparison: $14.93 for Claude Sonnet 5.5 and $11.62 for GPT-6 Astra. AA-3D measures preference for the resulting interactive scenes under this setup; it does not establish general model capability or production readiness.