Skyfall AI's SKY-001 world model matches GPT-6 Astra CAD accuracy at 150x lower cost
rohanpaul_ai · x · 2026-11-03
Skyfall AI released SKY-001, an engineering world model that designs physical products by planning parts and build steps in 3D — demoed designing a Jaguar car with thousands of parts.
- Outputs a buildable sequence: part tree, sketch per operation, and each intermediate solid
- Two stages: a Hierarchical Composer generates the object and unfolds it into assemblies/parts, then a Construction World Model predicts the visual build steps per leaf part
- Benchmark chart: on a 1,000-component CAD construction benchmark, SKY-001 matched GPT-6 Astra accuracy at $0.005 vs $0.75 per design (150x cheaper) and 3.9s vs 199s-1250s generation time; Grok 4.7 completed none
- The team argues world models are the next frontier after LLMs: hierarchical decomposition, constraint satisfaction and sequential construction directly in the physical domain
More from Models
- GDPx and Epoch's ECI leaderboards nearly agree, Spearman ρ=0.983 across 17 models — scaling01 · 2026-11-03
- TMax v1.1 released: open terminal agents under 30B gain up to 6 points on TB2 — natolambert · 2026-11-03
- DeepSeek's super-agent is 1.9x faster — but burns 3.8x the tokens — AIFlow_ML · 2026-11-03
- Team Replicates Anthropic's "Teaching Claude Why" Alignment Method on Qwen and Open-Sources It — austinc3301 · 2026-11-03
- User complains Sol 6.1 turned lazy: useless test loops, missing details, no progress — sloppenheimer · 2026-11-03
- 'Astra can rewrite Qwen weights' hits 2863 points on HN, then a rebuttal calls it misleading — yoshiyama_akira · 2026-11-03