October 13, 2026
Four models statistically tied at the top of AA-3D: leader Claude Sonnet 5.5 at 1,194 Elo
Artificial Analysis launched the AA-3D benchmark on Oct 12: 55 briefs, more than 10,000 counted human votes, 21 models. Four models are neck and neck at the top: Claude Sonnet 5.5 (1194 Elo), Claude Opus 5.5 (1190), GPT-6 Astra (1165) and Claude Fable 5.1 (1131).
Artificial Analysis
@artificialanlys
We measured intelligence. Now it's time to see what it can do. Launching AA-3D: our new benchmark for how AI models build interactive 3D scenes. Each model works as an agent in an offline sandbox and turns a written description into a three.js web page. A live human jury compares scenes built from the same description in pairs, without knowing which model made which. AA-3D details: ➤ 55 written briefs in seven areas: landmarks, vehicles, living organisms, games and space. A brief says what the scene should show and how to interact with it, and leaves the rest to the model. ➤ Each model gets up to three hours and 250 turns. Its sandbox has Blender, Rapier physics, Python, a headless browser to check its work, and a library of CC0-licensed materials, 3D models and characters. ➤ Jurors look at each scene for at least 30 seconds before voting. One in four comparisons is a check. ➤ More than 10,000 counted votes from more than 1,000 jurors rank 21 models by Elo, with 95% confidence intervals. Key findings: ➤ Four models are statistically tied at the top: Claude Sonnet 5.5 (1194), Claude Opus 5.5 (1190), GPT-6 Astra (1165) and Claude Fable 5.1 (1131). The confidence intervals of the other three overlap with Claude Sonnet 5.5's. ➤ Elo on AA-3D rises with intelligence: each point on the Artificial Analysis Intelligence Index is worth about 18 Elo (Pearson correlation 0.92). ➤ The top open-weights models, Kimi K3 (982) and MiMo-V2.6-Pro (977), rank 11th and 12th, more than 200 points behind first place. ➤ A scene costs from $0.20 to $32.08 depending on the model. The cheapest, GPT-6 Luna, ranks eighth. More results, scenes and a description of how AA-3D works will follow in the thread below.
· 30.4K views
Artificial Analysis already measures model intelligence with an index. Now it checks what models can build. A model works as an agent in an offline sandbox: it gets a written brief and builds a web page with three.js. A human jury blindly compares pairs of scenes built from the same brief.
Key point. The ranking rises with the intelligence index: each index point is worth about 18 Elo (Pearson correlation 0.92). The best open model, Kimi K3 (982), ranks 11th, more than 200 points behind the leader.
A scene costs from $0.20 to $32.08 depending on the model. The cheapest, GPT-6 Luna, ranks eighth.
More results, scenes and a description of the methodology are promised in the following posts of the thread.