Artificial Analysis adds a 3D-building benchmark, with four AI models tied at the top

Artificial Analysis released AA-3D on October 12, 2026, judging AI models by interactive 3D scenes people can explore—not just the code they write. The benchmark compares 21 models. Claude Sonnet 5.5 has the highest launch score, but four models are statistically tied at the top, and the cheapest scene builder ranks eighth.

A written brief becomes something to explore

Each model works as an agent in an offline sandbox, turning a written brief into a three.js web page. It can write code, render its scene, inspect screenshots and revise the result. The shared environment includes Blender, Rapier physics, Python, a browser without a visible window and a library of CC0 assets.

Every task comes with the same delivery contract. A run ends after three hours or 250 agent turns, leaving an index.html page that runs in a browser. The scenes span seven domains, including machines, living bodies, games and space. GPT-6 Luna, for example, completed all 55 scenes.

A recruited panel compares two scenes made from the same brief without knowing which model built either one. Artificial Analysis converts those preferences into Elo ratings, with higher scores indicating stronger results in the comparisons. The launch article reports more than 10,000 panel votes.

Sonnet 5.5 scored 1194. Artificial Analysis treats the top four as statistically tied because their 95% confidence intervals overlap with Sonnet’s; all four come from Anthropic and OpenAI. The highest-ranked models with openly available weights, Kimi K3 and MiMo-V2.6-Pro, finished 11th and 12th at 982 and 977.

Looking sooner tracks stronger scenes

The benchmark also records how agents work. The top five models first inspected an image of their own scene after a median of 10 to 13 turns. They did so by turn 20 in at least 89% of their scenes. Models ranked 14th through 20th waited until a median of 15 to 40 turns.

Across the 19 models able to view images, early inspection had a strong relationship with Elo: a Spearman rank correlation of 0.82. That compares behavior across models; it does not demonstrate that moving an agent’s first inspection earlier would improve its score.

Frequency told a different story. MiniMax-M3 inspected its own scenes 54 times per scene, more than any other model. Yet it reached its first inspection by turn 20 in only 7% of scenes and ranked 19th. Neither output-token volume nor turn count showed a statistically significant relationship with Elo.

More detailed scenes won more votes, but much of that gap reflected which model built them. Within each task, the most detailed fifth won 63% of votes, versus 31% for the least detailed fifth. Within one model’s outputs, a scene one standard deviation more detailed won only about 1.5 percentage points more votes.

The highest score is not the lowest bill

Input accounted for 69% of spending because agents resend their accumulated context on every turn. That makes the bill about more than the code produced at the end: repeated processing of the agent’s working history accounts for most of the reported spend.

• Among the top five, GPT-6 Sol had the lowest mean cost at $3.46 per scene. Sonnet 5.5 cost 4.4 times as much, and Fable 5.1 cost 9.3 times as much.
• GPT-6 Luna ranked eighth at $0.20 per scene. Kimi K3 spent 68 times as much per scene for a similar Elo rating.

GPT-6 Sol was also the quickest of the top five across the six tasks shown in the run comparison. Its median run took 60 minutes, against 115 to 151 minutes for the other four. Price, elapsed time and preference rank therefore offer different ways to compare these scene builders.

Readers can explore the generated scenes and replay agent runs in the launch article. Its public voting exercise reveals the model after a choice, but those votes do not affect the leaderboard. Artificial Analysis plans to add models as they are released.