Frontier3DBench evaluates 11 coding agent configurations for 3D world generation

·2026.11.09 12:41

Key point

The benchmark tests 145 tasks across 9 categories, finding that GPT-6-astra leads in interactive worlds while local mechanisms outperform stateful workflows.

Details

Einsia released Frontier3DBench to evaluate coding agents on building 3D objects, scenes, and interactive worlds.
GPT-6.1-sol leads overall and in object and scene tasks, while GPT-6-astra leads in interactive worlds.
MiniMax-M3.1-Flash ranks sixth overall but second in scene editing.
All configurations score lower on reference-based reconstruction than text-driven generation.
Local mechanisms score 83.08 out of 100 compared to 30.26 for stateful workflows across 165 world submissions.
The benchmark includes 145 tasks across 9 categories and 11 model-harness configurations.
Evaluation covers appearance, structure, and behavior via rendered views, scene inspection, and runtime tests.

benchmark 3d-generation unreal-engine coding-agents gpt-6.1-sol einsia

Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.