Artificial Analysis launches AA-3D benchmark for AI models in 3D scene generation

AA-3D: AA-3D is Artificial Analysis's benchmark for assessing AI models' ability to generate interactive 3D web experiences using three.js. Models act as agents with access to tools like Blender, physics engines, and a headless browser to iteratively create scenes from detailed text briefs across domains such as landmarks, machines, and living bodies. Results are determined through large-scale blind pairwise comparisons by human judges converted to Elo ratings. Kimi K3: Kimi K3 is an open-weights AI model that competed in the AA-3D benchmark. It achieved a mid-tier ranking among the 21 evaluated models in generating interactive 3D scenes. Its results highlighted the performance gap between leading closed models and open-weights alternatives in this agentic 3D task. GPT-6 Sol: GPT-6 Sol is an OpenAI model that ranked highly in the AA-3D benchmark and appeared on the cost frontier. It demonstrated effective agentic behavior in constructing interactive 3D experiences from briefs. The model showed strong domain performance alongside other top entries. GPT-6 Luna: GPT-6 Luna is an AI model from OpenAI that participated in the AA-3D evaluation and placed on the cost-performance frontier. It produced functional interactive 3D scenes while operating at a lower cost point than many higher-ranked competitors. Its results contributed to insights on efficient model performance in 3D generation tasks. GPT-6 Astra: GPT-6 Astra is an advanced AI model associated with OpenAI that achieved a top-tier ranking in the AA-3D benchmark. It successfully generated competitive interactive 3D scenes from written briefs using iterative agentic processes. The model was statistically tied with the leading entries based on human panel votes. MiMo-V2.6-Pro: MiMo-V2.6-Pro is an open-weights AI model evaluated in the AA-3D benchmark. It placed in the middle of the leaderboard for its ability to build interactive 3D web scenes from text descriptions. The model illustrated the current standing of open models relative to frontier closed systems in complex creative coding tasks. Claude Opus 5.5: Claude Opus 5.5 is a high-capability AI model developed by Anthropic that placed near the top of the AA-3D leaderboard. It excelled at building detailed interactive 3D environments from textual instructions within the benchmark's sandbox constraints. Its performance aligned closely with other top models in blind evaluations across multiple domains. Claude Sonnet 5.5: Claude Sonnet 5.5 is a frontier AI model from Anthropic that ranked among the top performers in the AA-3D benchmark. It demonstrated strong capabilities in translating complex briefs into functional interactive 3D scenes through agentic workflows. The model tied statistically with other leading entries in overall preference from human judges. Artificial Analysis: Artificial Analysis develops and maintains a suite of benchmarks that evaluate AI models across capabilities like reasoning, coding, and agentic tasks. It created and launched the AA-3D benchmark to measure how effectively frontier models can autonomously build interactive 3D scenes from written briefs inside a controlled sandbox environment. The organization aggregates results from blind human voting panels to produce ranked leaderboards. AI Benchmarking: Human preference voting through blind pairwise comparisons is emerging as a reliable method for ranking AI models on creative and agentic tasks where objective metrics fall short. Agentic AI Development: Leading models increasingly incorporate early self-inspection of outputs and tool use like browsers to iteratively refine complex generations such as interactive 3D scenes. Model Performance Trends: Performance on interactive 3D scene building correlates strongly with broader intelligence measures, indicating that general advances in AI reasoning benefit specialized creative applications.