Independent benchmarking firm Artificial Analysis announced the AA-3D benchmark on 12 October 2026. The tool tests frontier AI models on their ability to generate functional interactive 3D scenes from text briefs by producing executable three.js code in a sandbox environment. Results are determined through blind pairwise preference votes by human judges using an Elo-style ranking system. This approach differs from existing benchmarks like T3Bench or Blender AI Arena, which focus on static 3D asset generation, by narrowing in on interactive three.js outputs. The human-preference format makes the benchmark harder to game, as judges represent subjective human judgment rather than fixed formulas, though panel composition could influence which models rank highest.
Source: Read the original article
