Artificial Analysis explains AA-3D evaluation method and publishes scene explorer

Artificial Analysis says its AA-3D evaluation has each model act as an agent in an offline sandbox to turn a written brief into a three.js web page, with up to three hours and 250 turns allowed.

Human panelists compare blind pairs of scenes built from the same brief, spending at least 30 seconds on each before voting. The votes are converted into Elo scores with 95% confidence intervals.

The company also published an interactive AA-3D Explorer where users can view every Claude Fable 5.2 scene. The explorer links to the full leaderboard and results for the evaluation.