Hobby Blender benchmark: GPT-6-Astra one-shot scenes strikingly outperform other tested models

Gruku · reddit · 2026-09-08

A developer running a personal Blender benchmark (inspired by MineBench) tested GPT-6-Astra on one-shot Python script scene generation, finding results strikingly better than all other models tried. Blender MCP multi-turn runs looked even more impressive but were too costly to test at scale. The author feels one-shot "make something visually interesting" is hitting its ceiling; next up are game-ready assets and real production tasks. Results and voting at blenderbench.realityreprojector.com.

Original post →

More from Models

Models channel →