Skyfall launches engineering world model: CAD part generation 51 times faster than GPT-6 Astra, at just 1/150 the cost

Beating AI News Flash: U.S. AI startup Skyfall AI has released an engineering world model, SKY-001, specifically designed to handle mechanical design and architectural floor plans. The model took 18 months to develop. Founder Sam Pasupalak previously co-founded deep learning company Maluuba, which was acquired by Microsoft in 2017.

SKY-001 does more than generate attractive 3D appearances; it also attempts to reconstruct the internal structure of objects and plan the modeling steps for parts. For example, given a photo of a car, it first generates the rough shape of the entire vehicle, then breaks it down layer by layer into different components, generating sketches and 3D modeling processes for each part. In architectural design, it can arrange apartments, corridors, and interior rooms based on floor outlines.

Its core approach is to plan the overall structure first, then break complex designs into many small tasks so that different parts can be generated simultaneously. When drawing sketches, it also first draws the rough outline and then fills in the details. Compared with having a language model generate step by step along a long chain of operations, this method can reduce redundant computation and prevent errors from accumulating.

Skyfall's published test of 1,000 CAD parts shows that SKY-001 takes an average of 3.9 seconds per case, while GPT-6 Astra requires 199 seconds, a 51-fold difference in speed. The officially estimated cost per case is $0.005 and $0.75, respectively, a 150-fold difference. Their geometric overlap is 79.7% and 79.1%, respectively, so the generation accuracy is actually not much different.

In architectural floor plan tests, SKY-001 also leads GPT-6 Astra, reducing generation time from 144.6 seconds to 1.22 seconds, about 119 times faster, while geometric overlap increases from 40.6% to 43.6%.

Currently, SKY-001 is still in the research preview stage and has not yet released an API or model weights. It can already demonstrate how parts are built step by step, but these modeling steps cannot yet be directly handed to CAD software for execution and modification.