What changed
Artificial Analysis introduced AA-3D on October 12. Models receive a written brief, build an interactive three.js scene in an offline sandbox, and can inspect and revise their work. A recruited panel then compares pairs of scenes without knowing which model created each one.
The useful lesson
The launch report lists a highest score, but says the leading four models are statistically tied. That distinction matters: a table’s first row can look more decisive than the evidence supports. The article also separates the launch snapshot from the live leaderboard.
Try reading it this way
Look at the scenes and decide what you value: clarity, useful controls, visual detail, or responsiveness. Then read the setup and limits. We have not reproduced this evaluation, and a preference test for 3D scenes is not a universal measure of an assistant.