Artificial Analysis has a new way to test frontier AI models: ask them to build a 3D world, then let humans decide whose world is better.
The independent AI benchmarking firm announced AA-3D on 12 October 2026. The benchmark measures how well models can generate interactive 3D scenes from written briefs, with results ranked by human judges using an Elo-style rating system.
How AA-3D works
A model receives a text brief describing a scene. It then has to produce a functional three.js experience from that description.
For the uninitiated, three.js is a JavaScript library used to render 3D graphics in a web browser. The model is not drawing a picture of a scene. It is writing the code that builds one.
Those outputs are then run in an offline sandbox environment.
A recruited panel casts blind pairwise preference votes. Judges see two outputs side by side, without knowing which model made which, and pick the one they prefer.
Those head-to-head results feed an Elo-style ranking.
A familiar playbook from Artificial Analysis
AA-3D is not the firm’s first experiment with human preference as a yardstick. The most recent example is AA-Video-T2V v2.0, which launched around 30 September 2026. That text-to-video benchmark drew more than 68,000 votes across 1,000 prompts.
The new benchmark also stakes out territory that existing efforts do not fully cover. Academic and community projects such as T3Bench and Blender AI Arena already assess text-to-3D tasks. Their focus, however, is broader text-to-3D generation and static 3D assets.
AA-3D narrows the lens to interactive three.js outputs.
Why interactive 3D is a tougher test
Generating a convincing 3D model is already a demanding task. Generating one that actually works as interactive software raises the bar in several ways at once.
The model has to understand the brief, translate that brief into code that runs without breaking, and produce a result that looks good and feels coherent once a human starts interacting with it.
That combination makes AA-3D something of a hybrid exam. It blends coding ability, spatial reasoning and visual judgment into a single score. Traditional coding benchmarks often check whether code passes tests. This one checks whether humans actually like what the code produces.
What this means for AI developers and users
The human-preference format matters here. Automated metrics can sometimes be gamed by tuning a model toward whatever the scoring system rewards. Blind pairwise voting is harder to optimize against, because the target is subjective human judgment rather than a fixed formula.
That does not make the approach flawless. Preference votes reflect the tastes of whoever sits on the panel. How the panel is assembled could influence which models rise to the top.
As of the announcement, there are no publicly available details on the methodology, prompt set size, or results for AA-3D.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.