Skyfall AI, led by former Maluuba CEO Sam Pasupalak, has unveiled SKY-001, a model it calls the first Engineering World Model. The pitch is blunt: on CAD and BIM tasks it matches or beats GPT-6 Astra and other frontier models while running roughly 50x to 120x faster and costing about 150x to 185x less per case.
That is a big claim, and the replies on X ran far ahead of it. This post walks through how SKY-001 works, reads the benchmark tables line by line, and marks what is claimed versus what anyone outside the company has checked. For background on the category, start with our explainer on world models and the Starchild-1 launch.
TL;DR: SKY-001 at a glance
| Question | Short answer |
|---|---|
| Who made it? | Skyfall AI, an "AI lab advancing intelligence through world models," led by Sam Pasupalak |
| Announced? | October 19, 2026, with a research post titled "Introducing SKY-001: The First World Model for Engineering" |
| What does it do? | Image or text to a tree of CAD parts and their build steps; building outline to a floor plan, units and rooms |
| CAD result vs GPT-6 Astra | 79.7% vs 79.1% IoU, 3.9 s vs 199 s, $0.005 vs $0.75 per case |
| BIM result vs GPT-6 Astra | 43.6% vs 40.6% IoU, 1.22 s vs 144.6 s, $0.0024 vs $0.446 per case |
| Can I use it? | Waitlist only; no public weights or price list yet |
| Who checked the numbers? | Nobody outside Skyfall, as far as we can find |
| Biggest caveat | The benchmarks are Skyfall's own, and the accuracy lead is small; the real gap is speed and cost |
What is Skyfall actually claiming?
The launch thread on X says SKY-001 "beats GPT-6 Astra and every frontier model we tested" on CAD and BIM, at 51x to 125x faster and 150x to 210x cheaper, with a higher IoU than all of them. IoU, or intersection over union, scores how much a generated shape overlaps the reference shape. A score of 100% would be a perfect match.
The thread also argues a thesis: language models are autoregressive, so they predict one token at a time, and were never built for visual reasoning or long-horizon tasks. Skyfall says it compared SKY-001 against GPT-6 Astra, Claude Opus 5.5 and others.
The accompanying research post is more careful than the tweets. It says SKY-001 "matches or exceeds frontier models in a fraction of the time and cost." That is a different sentence from "beats every frontier model," and the tables back the careful version.
How does a world model design a physical object?
A flat language model writes CAD code in one long chain, where every decision depends on all the others and errors pile up over thousands of steps. Skyfall's answer is hierarchy. SKY-001 has two stages, and both work in pictures instead of text.
Branching design tree with one highlighted path, a picture of the hierarchical composer splitting an engineering world model design into assemblies and parts
Stage 1: the Hierarchical Composer
The composer decides what a design is made of. Given an image, text, or spatial constraints, it first generates a rough whole, then splits it level by level into assemblies and parts. Once a parent is fixed, its children can be generated in parallel, "the way an engineering team divides the work."
Skyfall's key argument is about depth. The number of sequential steps grows with the depth of the design rather than its size, so a decision about a single bolt can stay consistent with a whole engine. In the post's example, a VMS-85 engine photographed once becomes a five-level tree of parts. A refiner then fills in what the rough shape lacked: hidden interiors, gear stacks behind covers, sleeves buried in housings.
Stage 2: the Construction World Model
The second stage builds each leaf as an ordered sequence of visual states. In CAD, it plans the intermediate 3D states of a part and draws the 2D sketch that turns one state into the next, such as a profile becoming an extrusion, then a cut that makes a hole. In BIM, it draws the floor plan one state at a time and decides for itself when the drawing is done.
Under the hood both use diffusion in two passes. A coarse rollout runs in a compressed latent space built by a Matryoshka-style autoencoder, so the earliest codes carry the global layout. A refiner then redraws each state at full resolution.
What does the benchmark table really show?
The numbers below come from Skyfall's own post, not an independent lab. Each row is a per-case mean.
Coin stack shrinking step by step, a picture of the cost drop Skyfall claims for SKY-001 CAD generation against frontier models
CAD construction (1,000 components)
The benchmark samples 1,000 components from complex objects such as a Jaguar E-Type Roadster, a FANUC ArcMate 100iB robot arm, a Minitractor 4x4, a Bell 222 helicopter and a jet engine. Each takes about 100 drawing steps and is scored by IoU against the reference.
| Model | Time per case | Cost per case | Mean IoU |
|---|---|---|---|
| SKY-001 | 3.9 s | $0.005 | 79.7% |
| GPT-6 Astra | 199 s | $0.75 | 79.1% |
| Gemini 4 | 253 s | $0.25 | 76.9% |
| Qwen 3.8 Max | 764 s | $0.30 | 34.4% |
| Kimi K3 | 439 s | $0.13 | 9.5% |
| Grok 4.7 | 1,250 s | $0.59 | no output |
Do the arithmetic and the headline holds: 199 divided by 3.9 is about 51x, and 0.75 divided by 0.005 is 150x. But the accuracy line is a 0.6-point edge over GPT-6 Astra. Skyfall itself calls this "matching its performance." Against the cheaper Gemini 4, the cost advantage shrinks to about 50x.
BIM floor plans (1,000 plans)
The BIM benchmark mixes about 350 whole floors from IFC building models and 650 apartment-level plans, at roughly 70 drawing steps on average.
| Model | Time per case | Cost per case | Mean IoU |
|---|---|---|---|
| SKY-001 | 1.22 s | $0.0024 | 43.6% |
| GPT-6 Astra | 144.6 s | $0.446 | 40.6% |
| Gemini 3.8 Flash | 346.2 s | $0.218 | 42.3% |
| Qwen 3.8 Max | 255.3 s | $0.111 | 28.8% |
| Kimi K3 | 113.9 s | $0.270 | 33.7% |
| Grok 4.7 | 592.1 s | $0.305 | 28.2% |
Here SKY-001 is about 119x faster and 186x cheaper than GPT-6 Astra, and leads on IoU by 3 points. Notice the absolute level, though: 43.6% means the best result still overlaps the reference less than half the time on average. That is a useful lead, not a solved task.
The composer judged by other models
For the image-to-parts composer, Skyfall used two blind LLM judges, GPT-6.1 Sol and Claude Opus 5.5, scoring renders from SKY-001 and GPT-6 Astra on a 1 to 10 scale. SKY-001 scored 7.8 on part decomposition against 6.4, 7.3 against 6.3 on interior, and 8.1 against 6.7 on match to the input image. Exterior realism was close at 7.7 against 7.4. The post notes GPT-6 Astra returned a model for only 82% of assets, which affects how that comparison reads.
Claimed vs verified: what we can and cannot confirm
We read the research post and the X thread. We have not run SKY-001, because access is by waitlist, so nothing below is hands-on testing.
| Claim | Status |
|---|---|
| SKY-001 is far faster and cheaper per case than frontier LLMs | Consistent with Skyfall's own tables; plausible for a small specialized model against large general ones, but unreplicated |
| SKY-001 beats every frontier model on IoU | Overstated in the tweet; the post says "matches or exceeds," and the CAD lead is 0.6 points |
| It is a "better engineering AI than GPT-6 Astra" overall | A reply-guy summary, not a Skyfall claim; benchmarks cover two task families only |
| Training data is far richer than public datasets | Skyfall reports a median of 288 parts per object against 6 in PartVerse; self-reported |
| Results transfer to real manufacturing | Not tested in the post; outputs are sketches and states, with executable feature histories listed as future work |
Two framing points matter. First, the benchmark design is Skyfall's, including the choice of competitors, prompts and how a language model is asked to write CAD code. Second, the post itself says its sketch and 3D transitions are "illustrative" and that the saved sketch and 3D geometry can disagree.
One amplifying post on X was labeled a paid partnership. The label does not say anything about the model, but it is a reason to weigh enthusiasm in the replies as marketing until independent tests exist.
Why does this matter if you build with AI?
The interesting point is not the leaderboard line. It is the cost structure. If a specialized visual world model can produce a comparable CAD result for half a cent instead of 75 cents, then the economics of agentic design loops change. A loop that tries 200 candidate designs costs about a dollar instead of $150.
That sets up a division of labor that other recent releases hint at: a general LLM handles instructions and tool calls, and a small specialist handles the geometry. We have seen the same shape of argument in Liquid AI's D1 decision model priced at 200x below GPT-6.1 Sol and in the claimed-vs-verified read on Elorian's Glimpse. Each says a narrow model can undercut a frontier model on one job. Each deserves the same caution until someone else runs the test.
Crane lifting a block onto a stack, a picture of construction sequencing in Skyfall SKY-001 BIM floor plan generation
How does it compare with other CAD and world-model work?
If you want to generate CAD today, the practical options are code-first. Our guides to SolidCraft, an open-source Rust parametric CAD server for agents and the text-to-CAD agent plugin that emits STEP and STL let a coding agent write geometry through tools you can run now. SKY-001 takes the opposite bet, planning in pictures with a trained model you cannot yet touch.
On the world-model side, SKY-001 is a different animal from the robot models we have covered. GreenWAM and NVIDIA's Long-WAM predict how a scene changes when a robot acts. Dream Labs' Dream-1 sells robot control as an API. SKY-001 predicts how a design gets built. For another 3D benchmark where frontier models struggle, see Qwen's BlenderWorld.
What people are asking
Is this the end of CAD jobs? Nothing in the post says so. Skyfall lists editable feature histories, text instructions such as "widen this bracket," and executable operations as next steps, which means the current output is a visual plan, not an editable model an engineer can open in their tool.
Why compare against language models at all? Because engineers now ask coding agents to write CAD and Blender scripts. Skyfall is arguing the right tool is a visual model, and the cost gap is the proof it offers.
Did the Indian-founder framing matter? One widely shared reply praised Pasupalak, who grew up in Odisha, and called for India to build dedicated innovation towns. That is commentary on the ecosystem, not a fact about the model.
What is the long-term plan? The post closes by pointing at PCB design and other hardware, "as a step toward Skyfall's longer-term vision of autonomous businesses." Pasupalak describes the company as building enterprise superintelligence.
What to do next
If you work in CAD, BIM or manufacturing tooling, join the waitlist and ask for the evaluation harness, not just a demo. Specifically ask for three things:
- The prompts and scripts used for each frontier-model baseline.
- IoU on your own parts, not the benchmark's parts.
- Whether outputs export to an editable feature tree.
If you build agents, treat the cost gap as a design hint. Route geometry to a specialist model if one proves out, and keep a general LLM for planning and tool use. Watch for a third-party benchmark before you change a roadmap. For help choosing models by task and price, our GPT-6.1 Sol and Astra cost-efficiency breakdown shows the baseline numbers SKY-001 is being measured against.
Related reading
- What are world models? Starchild-1 and Odyssey explained
- GreenWAM: open robot world model for 20 robot families
- Long-WAM: NVIDIA world-action model
- Dream-1 robot world-action model API
- SolidCraft open-source parametric CAD for agents
- Text-to-CAD agent plugin guide
- Qwen BlenderWorld benchmark
- SKY-001 vs GPT-6 Astra vs Gemini 4 vs robot world models
- Official: Skyfall's SKY-001 research post
Figures, prices and benchmark numbers come from Skyfall AI's launch post and X thread of October 19, 2026, and are accurate as of the publication date. They are the company's own results and may change as independent tests appear.