Flying Into AI's 3D Benchmark

Artificial Analysis has released a new tool: AA-3D Explorer. This organization, known for AI model evaluation, has turned evaluation from “looking at scores” into “walking around the scene.”

Traditional comparison of 3D generation models is to pull up leaderboards and look at scores. But with something like 3D, no matter how high the score is, it’s better to see for yourself what the scene the model built actually looks like.

The idea behind AA-3D Explorer is straightforward: choose a task from 7 domains, and let different models each build a 3D scene from the same prompt. You can fly into it in the browser, move through the scene, and see how each model handles the same requirement. This is also consistent with their voting process: human review panels browse the scenes this way before voting.

The video above shows a human anatomy scene: a barbell squat movement with clear muscle annotations. For the same task, the differences between models are obvious just by looking at the scenes.

The model list is ordered by AA-3D Elo ranking; from top to bottom you can see all models’ performance on one task. Use the left and right arrow keys to switch tasks, and the up and down arrow keys to switch models. Click the Prompt button to view a summary of what each model was asked to build.

This is the result generated by another model on the same task. Same brief, two versions of the anatomy scene—the differences in detail are visible at a glance.

The tool also offers some practical features: you can take a screenshot to save the current view, and record up to 60 seconds of interaction. “3D only” mode hides model text and buttons, leaving only the pure 3D scene. Each scene has its own link, and sharing it includes a preview card.

The mountain bike scene is also among the 7 domains. A model’s ability to reproduce terrain, vegetation, and signage can be judged after taking a spin around the scene.

Some users say that this “same brief, switch models left and right” comparison method makes it obvious at a glance who is actually modeling and who is faking it with UI. Someone was still comparing 7 models in it at 2 a.m. Others say it feels like “flying into a benchmark.”

Currently in the AA-3D voting rankings, among the major labs, the highest real-user score is gemini 3.7 flash, ranking 7th out of 317 models (vibe score 7.2), followed by gpt 6.1 sol (16th) and opus 5.5 (23rd). Google’s latest gemini 4 argon currently ranks 53rd.

The leaderboard has become a 3D scene you can walk into—that’s the change AA-3D Explorer brings.