artificial analysis has introduced aa-3d, a benchmark that evaluates how well ai models build interactive 3d scenes from text descriptions using three.js. it tests 21 models across 55 tasks in seven categories. models can write code, inspect rendered scenes, and revise their work before submitting a final result.
aa-3d benchmark results
key benchmark results:
- claude sonnet 5.5 leads with 1194 elo, followed by claude opus 5.5 (1190) and gpt-6 astra (1165). the top four models are statistically tied;
- generation costs vary widely: from $0.20 per scene for gpt-6 luna, which ranks eighth, to $32.08 for claude fable 5.1;
- visual feedback matters: early inspection of generated scenes correlates with higher scores, while producing more tokens does not necessarily improve results.
the leaders are close in quality, while the cost gap between models reaches about 160x.
how it differs from other benchmarks
aa-3d evaluates how well a model can create a complete 3d scene in a browser, from writing code to testing and refining the result. blind human comparisons make it possible to assess the final output as a whole. publishing generation costs helps compare quality against cost. unlike 3dcodebench and 3jsbench, which combine human evaluations with automated checks, aa-3d relies more heavily on human preferences. this helps assess visual quality and interactivity. however, it makes it harder to objectively compare the technical accuracy of 3d scenes.
what aa-3d means for ai evaluation
aa-3d adds another dimension to ai evaluation by measuring the quality of the finished experience, not just whether generated code works. it could encourage more comprehensive comparisons of model quality, interaction, and cost, although its rankings remain specific to the tasks and environment tested.
