Blender Bench tests how LLMs build 3D scenes through Blender and MCP

Gruku · reddit · 2026-07-21

A hobby project called Blender Bench evaluates how well LLMs can build 3D scenes in Blender through MCP or scripted one-shot prompting.

The author standardizes renders so models can be compared fairly on the same scene-building task.
The setup deliberately excludes external 3D generators, focusing on LLMs directly controlling Blender.
Early runs mostly used GPT-5.6 family models because they were the only ones that could handle the task at all; the author says 5.6 Luna looks especially strong on price/performance, while 5.6 Sol Max is in another league.
Roughly $50 was spent already on a small subset of tasks, and the project is still in the “make it exist” stage before trying production work.
A public site lets others inspect and compare model outputs.

More from Multimodal