Atlas

Benchmarks

← All benchmarks

BlenderWorld

Code · 2026-10-14

223 real-world 3D creation tasks in Blender from Alibaba Token Hub and Zhejiang University, 78 specified in text and 145 with image or video references, across nine creative domains from architecture to characters and simulation. A coding agent works in an isolated GPU container with shell, file and image tools plus a Blender MCP server, and must deliver an editable .blend project with the requested renders or video. A GPT-5.6 Terra judge scores each task as the equal-weight mean of a weighted binary-rubric pass rate and eight 1-10 visual-dimension ratings, both on 0-100; a missing .blend scores zero. The score is the mean fused task score over all 223 tasks.

Top models (higher is better)

ModelScore
GPT-6 Astra79.9
Claude Opus 573.0
GPT-5.6 Sol67.7
DeepSeek V4.1 Flash65.7
GLM-5.3 Flash59.7
Opus 4.857.0
Grok 4.657.0
Qwen3.8 27B47.7
Kimi K346.5
Gemini 3.8 Flash34.2
Loading Atlas data…