Artificial Analysis Launches AA-3D Benchmark for Agentic Interactive WebGL Generation
Artificial Analysis introduced AA-3D, a benchmark evaluating frontier models generating interactive 3D web scenes using Three.js and Blender. Claude Sonnet 5.5 and Claude Opus 5.5 lead the leaderboard across 10,000 blind human evaluations.
Why it matters
Use the AA-3D sandbox tooling stack (Blender, Rapier physics, and headless browsers) as a proven reference architecture for vibe coding 3D web applications.
TL;DR
- 01Claude Sonnet 5.5 and Claude Opus 5.5 lead interactive 3D WebGL generation, tied statistically with GPT-6 Astra.
- 02Open-weight models currently lag frontier closed models by over 200 Elo points in 3D agent tasks.
- 03Effective 3D agent sandboxes require three core tools: headless browser feedback, Blender scripting, and a physics engine like Rapier.
Key facts
- Evaluated Models
- 21 models
- Top Elo Score (Claude Sonnet 5.5)
- 1194 Elo
- Runner-Up Elo Score (Claude Opus 5.5)
- 1190 Elo
- Top Open-Weights Model (Kimi K3)
- 982 Elo
- Cost per Scene Generation
- $0.20 to $32.08
- Agent Execution Limit
- 250 turns, 3 hours
The AA-3D Evaluation Methodology
The AA-3D benchmark challenges LLMs with end-to-end frontend 3D development rather than simple script output. Each evaluated model is deployed as an agent within an isolated sandbox containing:
- Asset and Geometry Tooling: Blender and Python for programmatic 3D asset manipulation, alongside a CC0 asset library of materials and models.
- Physics Engine: Rapier physics for interactive rigid-body and collision simulation in JavaScript.
- Self-Verification Loop: A headless browser interface allowing the model to capture screenshots and verify rendering issues across up to 250 agent turns.
- Evaluation Constraints: A runtime budget of up to 3 hours per scene across 55 standardized design briefs.
Leaderboard Results and Open Model Gap
More than 1,000 human panelists reviewed paired scenes blindly, logging over 10,000 votes. Four frontier models clustered at the top within overlapping 95% confidence intervals:
1. Claude Sonnet 5.5: 1194 Elo 2. Claude Opus 5.5: 1190 Elo 3. GPT-6 Astra: 1165 Elo 4. Claude Fable 5.1: 1131 Elo
Open-weights models lagged significantly: Kimi K3 ranked 11th with 982 Elo, and MiMo-V2.6-Pro ranked 12th with 977 Elo. The benchmark established a strong linear correlation between the Artificial Analysis Intelligence Index and AA-3D performance (Pearson correlation 0.92, approximately 18 Elo points per intelligence index point).
Cost Dynamics for Vibe Coders
Compute and token expenditure ranged between $0.20 and $32.08 per scene. The eighth-ranked model, GPT-6 Luna, delivered the lowest cost profile, providing a viable compromise for budget-constrained agentic pipelines.
Try it in 2 minutes
{
"agent_environment": {
"runtime": "node-22",
"headless_browser": "playwright-chromium",
"libraries": ["three", "@dimforge/rapier3d-compat"],
"tools": ["blender_headless_cli", "python3_script_runner"],
"max_turns": 250
}
}json
✓ When to use
- Vibe coding interactive web prototypes, browser games, and Three.js physics visualizers.
- Architecting autonomous visual agent sandboxes that require automated screenshot validation.
✕ When NOT to use
- Static 2D image asset generation where diffusion or direct rasterization models are required.
- Simple backend REST API scripting with no browser visualization or canvas requirements.
What to do today
- Equip local coding agent sandboxes with a headless Chromium instance to enable visual self-correction on WebGL and Three.js code.
- Bundle physics libraries like Rapier directly into agent prompt contexts when requesting complex canvas interactivity.
- Select GPT-6 Luna or Claude Sonnet 5.5 to balance iteration quality against API token costs when prototyping 3D web applications.