SKY-001 Is INSANE: 50x Faster Than GPT-6 Astra at CAD? (I Read The Whole Blog)

Author: Prompt Engineer 48 · Watch on YouTube

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: SKY-001 Is INSANE: 50x Faster Than GPT-6 Astra at CAD? (I Read The Whole Blog) · Channel: Prompt Engineer 48 · 09:52 · Watch on YouTube

Topics: Enterprise AI applications, AI evaluation and benchmarks, AI models

Summary

The video reviews Skyfall’s claims about Sky 001, a spatial world model that generates hierarchical designs and the visual construction steps used to build them. It explains the system’s two-stage approach, its CAD and BIM examples, and the data and benchmarks presented in the launch post. On Skyfall’s reported tests, Sky 001 is much faster and cheaper than the compared models, with similar CAD quality, but BIM scores remain modest. The presenter finds the hierarchy and some caveats compelling, while emphasizing that the results are company-reported and the model and code could not be independently tested.

Key points

Sky 001 generates a design as a hierarchy of assemblies and parts, along with the sequence of steps used to construct them.
Its hierarchical approach is intended to reduce long chains of dependent decisions and let work on child components proceed in parallel.
For CAD, the system turns a single image into a multi-level parts tree; a separate refiner adds hidden geometry and fine details to rough part shapes.
For BIM, it uses an outline and brief to arrange walls, circulation, and apartments, then generates the drawing sequence.
The construction world model plans intermediate 3D states and sketches between them; the blog cautions that these illustrative sketches may not match the saved geometry.
Skyfall reports more than 200,000 parts for the composer, over 600,000 construction examples, and about 35,000 BIM floor plans with around 2 million drawing states.
On the reported CAD construction benchmark, Sky 001 took 3.9 seconds per case and cost half a cent, with a mean IOU of 79.7%; GPT-6 Astra scored 79.1% and was reported to be about 50 times slower and 150 times more expensive.
On the BIM benchmark, Sky 001 took about 1.2 seconds per plan and scored 43.6% IOU; the presenter notes that no tested system exceeded 45%.
The presenter says the benchmark numbers and data are Skyfall’s own, the outputs are not yet editable CAD with feature history, and independent testing was not possible.

Chapters

Start	Chapter	Summary
00:00	Sky 001’s claims and hierarchical design	The presenter introduces Skyfall’s speed and cost claims and explains how Sky 001 plans designs as hierarchies of assemblies and parts.
02:09	CAD and BIM examples; construction sequences	The video walks through the engine and floor-plan examples, then describes how the model generates intermediate states and sketches to build designs.
04:17	Diffusion method and training data	The presenter outlines the two-pass drawing process and summarizes Skyfall’s CAD and BIM data preparation and dataset sizes.
05:57	Benchmark results	The video compares reported CAD speed, cost, and quality, reviews composer ratings, and covers the BIM benchmark’s results and limitations.
08:49	Roadmap and the presenter’s assessment	The presenter describes Skyfall’s planned features, weighs what seems convincing against what remains unverified, and concludes with questions for viewers.

Quotes

It generates a design plus the sequence of steps that builds it. — 01:02

The transitions are illustrative, and the saved sketch and the 3D geometry can disagree. — 03:44

A lead of 0.6 of a point isn’t a win, it’s a tie. — 07:07

Nobody is above 45%. — 08:49

This post has no model or code I can run. — 09:21

Concepts and entities

Sky 001, Skyfall, GPT-6 Astra, CAD, BIM, spatial world model, hierarchical composer, construction world model, PartNet, RunPod

Full transcript (original language)

00:00 One photo of a jet engine, one model, and out comes a full tree of parts, plus the sketches and operations that build each one. That’s what Skyfall says its new model, Sky 001, does. And the claim I can’t stop looking at about 50 times faster than GPT-6 Astra at about 150 times lower cost. Is that real? Let’s read the evidence together. >> [music] >> Hey, I’m Prompt Engineer. Today we go through the Sky 001 blog claim by claim,

00:30 and I’ll tell you what’s actually shown and what isn’t. Here’s the problem. Skyfall starts with language models can write code and reason, but physical engineering is geometry. A design that reads well in text may not fit together, may not be physically feasible, may not be manufacturable. You have to keep thousands of components consistent with each other, and language is a clumsy medium for that. So, their bet is to plan directly in the visual world. They call Sky 001 a spatial world model. And the twist is it isn’t generating for you

01:02 to look at or walk around in. It generates a design plus the sequence of steps that builds it. How does it work? Think about how a real engineering team works. Nobody designs an engine in one flat pass. You do the overall design, then the assemblies, then the individual parts and how to make them, each level guiding the next. Sky 001 copies that order into stages. First, the hierarchical composer. It generates the target design and unfolds it level by level into a tree. Second, the construction world model. It predicts the visual sequence that builds each leaf of that tree. Put together, you get

01:36 what they call a visual design plan. Why does the hierarchy matter? Because a flat model has to generate every part and every step in one long chain, and errors pile up over thousands of steps. With a hierarchy, once a parent is fixed, its children can be worked on in parallel, like a team splitting up the work. So, the number of sequential steps grows with the depth of that design, not its size. And every level hands an explicit sub goal to the level below. So, a decision about a single bolt stays consistent with the whole engine. That’s the core idea, and honestly, it’s a clean one. It can also give you options.

02:09 From one intent, it can generate several candidate designs first, so you pick one before it gets unfolded into parts. Let’s start with CAD. You give the composer one image. It first makes a rough 3D shape of the whole object, just good enough to decide what it’s made of. Then, it splits that into main assemblies, splits each assembly again, and keeps going until every piece is a single part. Their showcase is a VMS 85 engine from one photo, and it builds a five-level tree. Block and cylinder heads, the intake and air filter,

02:40 exhaust headers, the timing drive, and inside those housings, gear stacks, shafts, hoses, and sleeves. Now, there’s a catch, and they’re up-front about it. That rough shape has coarse surfaces and nothing inside, so a gear stack behind the timing cover comes out as an open shell. So, a refiner goes to each leaf, fills in the hidden geometry, and adds the fine detail. The page shows five parts before and after completion, and the difference is exactly what you’d hope for. Then, BIM, which is building design. Same trick. You give it what an

03:10 architect can write down the outline of a story, the thickness of every wall, and a short brief with room sizes from one outline and one brief. It places the walls and circulation and divides the floor into three apartments around a shared corridor, stairs, and lifts. About 29 m across in their example. Now, stage two, the construction world model, and this is the part I find genuinely interesting. CAD parts are not made in one stroke. A profile becomes an extrusion, a new sketch adds or removes material, a cut makes a hole. Each intermediate state is a sub goal, so Sky 001 plans the 3D states in between, and

03:44 for every pair of consecutive states, it draws the 2D sketch that turns one into the next. On the page, one part from that engine is built in five steps. Extrude the base outline, a second extrusion, then three cuts, and they add a caveat I really appreciated. The transitions are illustrative, and the saved sketch and the 3D geometry can disagree. That is an honest line to see in a launch post. In BIM, it learns the drawing order, exterior walls first, then the walls between apartments and the corridors, then room by room with doors, windows, and fixtures, and it decides for itself when the drawing is

04:17 done. Their claim, a whole floor rolls out in about 1 second. How does the drawing work technically? Diffusion in two passes. First, a coarse rollout in a compressed latent space built with a Matryoshka-style autoencoder. Training drops codes from the end of a long sequence while still demanding the same reconstruction. So, the early codes carry the global layout and the later ones carry fine detail. The coarse pass only uses the leading codes, so it’s rough, but it keeps the layout. Then, a refiner redraws every state at full resolution. Sharp lines, exact corners, small features. Plan with the cheap

04:52 thing, polish with the expensive thing. Now, data. This is where a lot of these projects quietly live or die. Their CAD data comes from public data sets of real product assemblies, but the raw files are messed, inconsistent assembly trees, overlapping faces, duplicated geometry, parts with no names. So, every asset goes through a cleanup that rebuilds the hierarchy to follow standard engineering practice. Where the construction history is missing, they recover it with engineering tools and LLM workflows and verify it before training. The total’s

05:23 more than 200,000 parts for the composer, over 600,000 construction examples for the world model, and for BIM, about 35,000 real and synthetic floor plans, around 2 million drawing states. They also compare against PartNet, a data set widely used for part decomposition, parts per object. PartNet has a median of six and a 99th percentile of 33. Theirs has a median of 288 and a 99th percentile of 4,180. ABMX bicycle in PartNet has eight parts. Their fat bike has 988. Worth remembering, that’s their own data

05:57 measured by their own team, so treat it as a claim until others can inspect it. Okay, performance. Everyone will scroll straight here, so let’s be careful. First, CAD construction. The benchmark is 1,000 components sampled from things like a Jaguar E-Type, a Fanuc robot arm, a mini tractor, a Bell 222 helicopter, and a jet engine. About 100 drawing steps each, and the final 3D object is scored by IOU against the reference shape. Time per case, Sky001 3.9 seconds, GPT-6 Astra 199, Gemini 4 253, Kimiko 3 439, Qwen 3.8 Max 764,

06:37 and Grok 4.7 1,250 seconds. And they page notes, it completed none. Cost per case, Sky001 half a cent, GPT-6 Astra 75 cents, Gemini 4 25 cents, Qwen 30 cents, Kimiko 13 cents, Grok 59 cents. And quality, mean IOU, Sky001 79.7%, GPT-6 Astra 79.1, Gemini 4 76.9, Qwen 34.4, Kimiko 9.5, Grok no output. So,

07:07 against the best model, that’s about 50 times faster, about 150 times cheaper, and basically tied on quality. A lead of 0.6 of a point isn’t a win, it’s a tie. And to their credit, they say matching the speed and cost gap is the story. For the composer, they ran a different test. The same image goes to Sky001 and three frontier models, each asked to write a Blender script for the whole assembly in one request. GPT-6 Astra came out far ahead of the other frontier models. So, they compare against it directly with two blind LLM judges

07:40 scoring from the same renders. Out of 10, part decomposition 7.8 against 6.4, interior 7.3 against 6.3, complexity 8 against 7.3, match to the input image 8.1 against 6.7, exterior realism 7.7 against 7.4. Notice where the gap is biggest, part decomposition and matching the input, exactly where hierarchy should help, and it’s smallest on how the outside looks. Two things to keep in mind though, GPT-6 Astra only returned a model for 82% of assets and LLM judges

08:15 scoring renders is a soft metric. Now, the BIM benchmark, 1,000 floor plans, about 70 drawing steps on average, Sky-001 builds each one in about 1.2 seconds at a fraction of a cent. GPT-6 Astra takes 144 seconds at 44 cents. That’s roughly 120 times faster and 185 times cheaper. And on IOU, 43.6% against 40.6, Gemini 3.8 flash is right behind it 42.3. But look at the absolute numbers. Nobody is above 45%. This is a hard task and none of these systems have

08:49 solved it, including this one. What’s next according to Skyfall? In CAD, sketch rollouts become executable operations with an editable feature history and you could ask to widen a bracket or move a shaft and the neighboring parts update. In BIM, floor plans become real walls, doors, and rooms that architects keep editing and beyond that, PCB design and other hardware as a step toward their longer-term vision of autonomous businesses. So, here’s my take. What convinces me? The hierarchy idea, the honesty about rough shapes and illustrative transitions, and benchmarks

09:21 of 1,000 cases each. What I can’t verify? This post has no model or code I can run. So, every number here is Skyfall’s on Skyfall’s benchmarks with Skyfall’s data. I didn’t test it and the outputs aren’t editable CAD with a feature history yet. They say that’s next. So, the real question is whether this turns into something engineers can edit and trust. Would you trust a model that plans in pictures instead of code? Tell me in the comments. I genuinely want to know. If you’re looking for cheap and reliable GPUs, I use RunPod

09:52 and the link is in the description. The blog link is in the description, too. If you want me to dig into the Metro Skeletons or into world models for hardware, say so in the comments. See you in the next one.