GPT-6 Astra hits 95.9% on BenchCAD, suggesting 3D intelligence's 'ceiling' was never fundamental
hanjie_chen · x · 2026-09-08
BenchCAD author Haozhe Zhang reports that GPT-6 Astra scored 95.9% on BenchCAD, his team's benchmark for executable 3D CAD reasoning — a new SOTA. Beyond the leaderboard, he argues this dismantles the strongest argument against MLLMs: that language-dominated pretraining and scarce aligned multimodal data impose a hard ceiling on 3D intelligence. With the right synthetic data, multimodal post-training, executable feedback and agentic tool use, once-unreachable capabilities can be systematically engineered. 3D, long seen as AI's hardest frontier, is rapidly becoming solvable.
More from Models
- Sander Dieleman on Diffusion Models, Typicality, and Why ML Breakthroughs Are Closer Than They Look — sedielem · 2026-09-08
- Astra on low reasoning beats Sol on high, and runs twice as fast, devs confirm — steipete · 2026-09-08
- Astra Driving Blender via Computer Use Could Be AI's Fourth Demand Wave — firstadopter · 2026-09-08
- GPT-6 Astra Tops Blueprint-Bench 2, Nearing Human 3D Spatial Understanding — steipete · 2026-09-08
- Same $10/$50 price tag, very different bills: Astra vs Fable 5.1 token efficiency and cache economics — nearsync43 · 2026-09-08
- davinci-002 shutting down September 28, ending the GPT-3 era — altryne · 2026-09-08