SKY-001: Skyfall AI's Engineering World Model Beats GPT-6 Astra on Cost by 150x — and on Accuracy by 0.6 Points

Skyfall AI announced SKY-001 on 19 October 2026, and the claim you will see repeated is that it outperforms GPT-6 Astra and Claude Opus 5.5 on CAD and BIM work. The number underneath that claim is smaller and more interesting than the claim itself. On the vendor's own CAD construction benchmark, SKY-001 scores 79.7% IoU against GPT-6 Astra's 79.1%. That is six-tenths of a point — the kind of gap you get from a different random seed, not from a new architecture. What is not close is everything around it: 3.9 seconds per case against 199, and $0.005 against $0.75. The accuracy is a tie. The efficiency is a rout. Skyfall's own materials make the second point loudly and, in one place, make the first point for you without meaning to. The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

There is a second thing to know before any of those numbers matter to you: SKY-001 is not something you can call. It is in research preview, behind a waitlist, with full commercial deployment targeted for early 2027. No API, no pricing, no weights, no parameter count, no licence. This piece is about what the launch actually establishes — how the benchmarks hold up when you read the table rather than the headline, what the model genuinely does differently, and what a team that needs this capability today should do instead of waiting.

What Skyfall actually shipped

Skyfall AI is a San Francisco company with a Toronto research hub, founded by the team behind Maluuba, the deep-learning lab Microsoft acquired in 2017. Sam Pasupalak is CEO; Kaheer Suleman and Sumit Pasupalak are co-founders. It emerged from stealth in July 2026 with an unusual pitch — enterprise world models, and eventually an AI-run company — backed by Fidelity, Touring Capital, M13 and iNovia. SKY-001 is its first model release and the first concrete artifact behind that thesis.

The pitch in one sentence: engineering design is not a language problem, so stop solving it with a language model. An engine breaks into hundreds of parts nested several levels deep, each constrained by its neighbours; a floor plan is walls, then units, then rooms, then doors and fixtures, drawn in an order that matters. A model that emits one long autoregressive chain of text has to hold all of that in a single context and gets worse as the chain lengthens. SKY-001 instead plans in the visual world, in a hierarchy, with each level handing an explicit subgoal to the level below.

Two components do it. The Hierarchical Composer takes a design intent — an image, text, or a building envelope with a brief — and generates the whole design, then splits it level by level, refining each leaf in parallel. The Construction World Model then predicts the ordered sequence of visual states that builds each leaf: in CAD, a chain of 3D solid states with the 2D sketch drawn between each consecutive pair; in BIM, a floor plan drawn stroke by stroke. The drawing itself runs as diffusion in two passes, a coarse rollout in a compressed latent space from a Matryoshka-style autoencoder, then a refiner that restores full resolution. Both are diffusion transformers; neither predicts tokens.

Nothing about that architecture is publicly sized. Skyfall disclosed its data, not its model: more than 200,000 parts for the composer and over 600,000 construction examples for the world model, plus roughly 35,000 real and synthetic floor plans representing about two million drawing states. The CAD data was rebuilt from public assembly datasets whose hierarchies were inconsistent and whose parts were unnamed, and where construction histories were missing, Skyfall reverse-engineered them with engineering tools and LLM-assisted workflows and verified the result before training on it. The data-scale argument is real: against PartVerse, the common part-decomposition dataset, the median object has 6 parts and the 99th percentile 33. Skyfall's median is 288 and its 99th percentile is 4,180.

Reading the CAD benchmark properly

The CAD construction benchmark is 1,000 components sampled from complex objects — a Jaguar E-Type Roadster, a FANUC ArcMate 100iB welding arm, a Minitractor 4×4, a Bell 222 helicopter, a jet engine — each taking roughly 100 drawing steps, scored by IoU against the reference shape. Five frontier models were run on it. Every number below is Skyfall's, produced by Skyfall, on models Skyfall drove itself.

• Time per case — SKY-001 3.9 s; GPT-6 Astra 199 s; Gemini 4 253 s; Kimi K3 439 s; Qwen 3.8 Max 764 s; Grok 4.7 1,250 s and completed none.

• Cost per case — SKY-001 $0.005; GPT-6 Astra $0.75; Gemini 4 $0.25; Qwen 3.8 Max $0.30; Kimi K3 $0.13; Grok 4.7 $0.59 for no output.

• Quality (mean IoU) — SKY-001 79.7%; GPT-6 Astra 79.1%; Gemini 4 76.9%; Qwen 3.8 Max 34.4%; Kimi K3 9.5%; Grok 4.7 no result.

That is the whole table, and it rewards a second read. The efficiency story is as strong as advertised: 51× faster and 150× cheaper than the best rival, which at a thousand designs a week is the difference between a $5 line item and a $750 one. The quality story is a draw. SKY-001, GPT-6 Astra and Gemini 4 all land within three points of each other at the top; the collapse below them — 34.4%, 9.5%, no output — says more about how badly naive LLM prompting fails at long construction sequences than it says about the winner. Grok 4.7 finishing zero of 1,000 cases is a reminder that these are long, brittle rollouts, and that the benchmark is measuring whether a model can stay coherent for 100 steps at all.

One caveat on the cost column: Skyfall reports a cost per case but not the accounting behind it. Whether that is API list price, rented GPU time, or amortised self-hosted inference changes the comparison substantially — particularly for models Skyfall ran itself rather than called. Treat the ratios as directionally right and the absolute dollars as unexplained.

The efficiency claim moves depending on which page you read

Skyfall's launch materials do not agree with each other, which is worth knowing if you plan to quote them. The launch blog headlines "about a 50× speedup and 150 times less" per CAD case, and its chart labels read 51× and 150×. The SKY-001 product page's headline reads "3.9 s vs 199 s, 49× faster" and "$0.005 vs $0.75, 143× lower cost". The press release says "100x faster" and "150x lower cost". Same model, same benchmark, four different speed multipliers across three documents published on the same day.

Arithmetic settles the CAD case: 199 divided by 3.9 is 51.0, and $0.75 divided by $0.005 is 150. The blog is right and the other two pages are rounded, stale, or computed against a different baseline — the press release's 100× is not reconcilable with any number on the site. This is not a scandal; launch pages are written by different people on different deadlines and nobody diffs them. It is a reason to cite the primary table, not the headline, and to expect the multipliers you see in coverage this week to be wrong in whichever direction the writer happened to read first.

The landing page also reports a second quality panel that does not match the blog. Its 3D CAD chart shows SKY-001 at 76.1% IoU against GPT-6 Astra at 73.4%, Claude Opus 5.5 at 61.8%, Gemini 3.8 Flash at 45.9%, Qwen 3.8 Max at 26.6% and Kimi K3 at 6.5%. The blog's construction benchmark shows 79.7% against 79.1%. Different task, different score, same model name on both — and the margin over Astra goes from 0.6 points to 2.7 depending on which one you pick up. If you are going to argue about this model's accuracy, name the benchmark with the number.

BIM: a thinner lead, and Gemini is right behind

The BIM benchmark is 1,000 floor plans — about 350 whole floors taken from IFC building models and 650 at apartment level — with 20 to more than 200 drawing steps each, around 70 on average, scored by IoU on the finished plan.

• Time per plan — SKY-001 1.22 s; Kimi K3 113.9 s; GPT-6 Astra 144.6 s; Qwen 3.8 Max 255.3 s; Gemini 3.8 Flash 346.2 s; Grok 4.7 592.1 s.

• Cost per plan — SKY-001 $0.0024; Qwen 3.8 Max $0.111; Gemini 3.8 Flash $0.218; Kimi K3 $0.270; Grok 4.7 $0.305; GPT-6 Astra $0.446.

• Quality (mean IoU) — SKY-001 43.6%; Gemini 3.8 Flash 42.3%; GPT-6 Astra 40.6%; Kimi K3 33.7%; Qwen 3.8 Max 28.8%; Grok 4.7 28.2%.

The 119× speed and 186× cost figures against Astra are accurate. But look at the quality column: Gemini 3.8 Flash is 1.3 points behind SKY-001, and Gemini 3.8 Flash costs 91× more but is still cheap in absolute terms at 22 cents a plan. The gap that matters for a firm drawing thousands of plans is real; the gap that matters for a firm drawing dozens is a rounding error against a workflow change.

And look at the absolute number. The best model on this benchmark reconstructs 43.6% of a floor plan by intersection-over-union. That is a research-preview quality bar, stated plainly by the vendor's own scoring, not a "replace the architect" bar. Read in that light, the launch's real claim is narrower and more defensible than the tweet: a purpose-built spatial model now plans long construction sequences at a speed that makes bulk generation and search practical, at roughly human-benchmark-level fidelity to the frontier LLMs it was compared against — not that it has solved building design.

What the tweet got wrong

The framing that SKY-001 "outperforms GPT-6 Astra and Opus 5.5 on CAD and BIM tasks" oversells one of those two comparisons. Claude Opus 5.5 does not appear in either construction benchmark. It appears twice elsewhere: in a qualitative video where four models each wrote a Blender script for a whole assembly in one request — where Skyfall itself says GPT-6 Astra was "far ahead" of the other frontier models — and as one of the two blind LLM judges scoring SKY-001 against Astra, alongside GPT-6.1 Sol. The landing page's 3D CAD panel does include Opus 5.5 at 61.8%, and that is the only place a number for it exists.

On BIM there is no Opus 5.5 row at all. There is also no Opus 5.5 row in the CAD construction table. So the honest version of the claim is: SKY-001 roughly ties GPT-6 Astra on accuracy at a fraction of the time and cost, and beats the frontier models that were run — which included Gemini, Qwen and Kimi — by margins that grow as the task gets longer. The Opus part of the sentence is inherited from a video demo and a judge seat, not from a benchmark.

The composer comparison deserves its own caveat, because it is the softer of the two results. Skyfall ran SKY-001 against GPT-6 Astra, Claude Opus 5.5 and Claude Fable 5.1 on the same input photographs, with every LLM writing a Blender script for the entire assembly in one request. Two blind LLM judges scored the renders from 1 to 10 on five dimensions, and SKY-001 took four of five — part decomposition 7.8 to 6.4, interior 7.3 to 6.3, complexity 8.0 to 7.3, match to the input image 8.1 to 6.7, exterior realism 7.7 to 7.4. The caveats are in Skyfall's own footnote: the judges are LLMs, the setup is one-shot prompting with no tooling or iteration for the rivals, and GPT-6 Astra returned a model for only 82% of assets — so on 18% of the set SKY-001 was competing against nothing. A single-request Blender script is a deliberately unflattering way to run a frontier LLM at 3D assembly. That does not make the result false; it makes it a comparison between a purpose-built pipeline and an off-the-shelf prompt, which is exactly what the construction benchmarks also are.

What this means if you actually build engineering software

The interesting claim is not the benchmark margin. It is that a hierarchical spatial world model can hold a 100-step construction plan together at all, across five levels of assembly and thousands of parts, and produce something that scores within a point of a frontier LLM on geometric overlap. If that holds up under independent testing, the practical consequences are about search rather than replacement: when a design rollout costs half a cent and four seconds instead of 75 cents and three minutes, you can generate a thousand variants and pick one. That is a different product from a model that produces one design slowly.

The parts Skyfall says come next point the same way. CAD sketches would become executable parametric operations with an editable feature history, so a designer could ask to widen a bracket and have the neighbouring parts that must change follow. BIM output would become walls, doors, windows and rooms in an editable building model rather than a drawing. Beyond that, PCB layout and other hardware. Those are roadmap statements, not shipped features, and the research preview contains none of them.

There are also the questions a model release normally answers and this one does not. No parameter count. No architecture paper. No weights, no licence, no self-hosting path. No context window, because it is not a chat model. No published evaluation harness, so no way to re-run the benchmark on your own parts. And no independent measurement: Artificial Analysis, LMArena and the usual third-party boards have no SKY-001 entry, which is unsurprising for a model that has been public for a day and cannot be called. Every figure in this piece is vendor-reported and unreproduced, and the vendor's own pages disagree with each other by up to 2× on the headline speed multiplier. Independent replication is the thing to wait for before the numbers become load-bearing in a build decision.

If you need this capability this quarter

You cannot have SKY-001. The product page offers a waitlist, the press release says full commercial deployment is targeted for early 2027, and the one integration named is a demonstration inside Autodesk Fusion rather than a shipped connector. Nothing here is a reason to plan a 2026 pipeline around it.

What you can have is the model it was benchmarked against. GPT-6 Astra is routed on OrcaRouter as openai/gpt-6-astra, at OpenAI's list prices of $10.00 per million input tokens and $50.00 per million output tokens with a 1.05M-token context window, passed through at 0% markup — so if OpenAI changes the rate, the rate here changes the same day rather than after a margin recalculation. It carries an Artificial Analysis Intelligence Index of 54.7, a Coding Index of 76.9 and an Agentic Index of 51.6, which is the independent number SKY-001 does not have.

That is also the sane way to hold an unproven model in view without betting a production path on it. When SKY-001 does reach an API — early 2027 on the current plan, and possibly through a partner before Skyfall sells direct — the test worth running is your own parts, not a 1,000-component benchmark. Until then, the routing pattern that survives a model you have not evaluated is the boring one: keep the call in one endpoint, put a fallback chain behind it, and let the router move traffic when a provider is slow or a model stops being the best answer. OrcaRouter's failover exists precisely because a new model should earn production traffic on measured results, and no one has measured SKY-001 outside Skyfall.

What to watch

Three things would turn this from a strong vendor demo into a decision. First, an independent evaluation on a public harness — the construction benchmark is Skyfall's own, defined by Skyfall, with no released task set, and the 0.6-point CAD margin is well inside the range that a different scoring convention could flip. Second, the commercial shape: whether early 2027 arrives as an API with a price per design, as a plugin inside CAD and BIM tools where the buyer already works, or as an enterprise licence sold to firms rather than developers. The Autodesk Fusion mention hints at the second, which would make SKY-001 a feature of the design software you already own rather than a model you call. Third, whether the hierarchy generalises past CAD and BIM to the PCB and hardware work Skyfall names next — that is the claim that the architecture, not the training data, is the advantage.

For now the accurate summary is narrow. Skyfall AI has shipped the first credible engineering world model, and the thing it actually wins on is speed and cost per design, by one to two orders of magnitude, with accuracy that ties the frontier rather than beating it. That is a genuinely new capability and a real strategic bet against the LLM approach to physical design. It is also, today, a waitlist page and a vendor benchmark — and the six-tenths of a point, not the 150×, is the number that will decide whether it is more than that.