Frontier3DBench Puts 11 Coding Agents Through 145 3D-Creation Tasks, GPT-6.1-sol Tops at 66/100

Tsinghua, Naver Lab and Einsia.AI release a unified benchmark of 145 tasks across object, scene and interactive-world 3D creation. GPT-6.1-sol leads overall at 66.02/100, while GPT-6-astra wins on interactive worlds. All 11 configurations trail text-driven generation on reference reconstruction, and generated worlds score 83 on local mechanisms but just 20.82 on motion clearance and 30.26 on stateful workflows — flagging coordinated activity and reference fidelity as the biggest open bottlenecks for 3D coding agents.