Artificial Analysis has introduced AA-3D, a new benchmark designed to evaluate how effectively AI models generate interactive 3D scenes from text descriptions. The system requires models to write functional code using the three.js JavaScript library to create browser-based 3D experiences.
The evaluation process relies on a panel of human judges who perform blind pairwise comparisons. Judges view two anonymous outputs side by side and select their preference, with final rankings calculated using an Elo-style rating system.
This initiative follows the release of the AA-Video-T2V v2.0 benchmark, which launched on September 30, 2026, and gathered over 68,000 votes across 1,000 prompts. Unlike existing projects such as T3Bench or Blender AI Arena, which focus on static assets, AA-3D emphasizes functional, interactive software.
The benchmark assesses a combination of coding proficiency, spatial reasoning, and visual quality. As of the announcement on October 12, 2026, the company has not yet disclosed specific details regarding the methodology, the size of the prompt set, or initial results.
