Apollo World vs SKY-001: When "World Model" Means Three Different Jobs

Apollo World and SKY-001 were announced fifteen days apart and have almost nothing in common except a label. Apollo World is Baidu's agent-driven simulator, disclosed on 3 November 2026, which generates camera and lidar observations of a virtual city so that a robotaxi stack can be trained and graded on complete trips. SKY-001 is Skyfall AI's engineering world model, announced on 19 October 2026, which predicts the ordered sequence of visual states by which a mechanical part gets built — a drawing-by-drawing construction plan. One rolls a traffic scene forward in time. The other rolls a manufacturing process forward in time. Both are correctly called world models, and that is exactly the problem with the phrase.

The term has quietly become a job title rather than a category. Three products now wear it: models that generate watchable environments, models that simulate an environment an agent can act inside, and models that predict the state transitions of a process. Apollo World is the second. SKY-001 is the third. Comparing them directly is like comparing a flight simulator to a assembly manual: both describe how something unfolds, and you would never use one for the other's job.

SKY-001 predicts a build, not a street

Skyfall AI was founded by the team behind Maluuba, the Canadian research lab Microsoft acquired in 2017, and its backers reportedly include Fidelity, Touring Capital, M13 and iNovia. SKY-001 was announced on 19 October 2026 as a research preview with a waitlist. There is no API, no pricing, no licence and no weights; full commercial deployment is targeted for early 2027.

The architecture has two stages, both diffusion transformers operating on visual states rather than text tokens.

• The Hierarchical Composer generates a whole design, then recursively splits it into assemblies, sub-assemblies and parts, with each level parallelisable. This is how the model avoids generating a single monolithic drawing of a complex object.

• The Construction World Model predicts the ordered sequence of visual states that builds each leaf part, using a two-pass diffusion scheme over a Matryoshka-style autoencoder: a coarse latent rollout first, then a full-resolution refiner.

The training data claims are large and specific: more than 200,000 parts for the composer, more than 600,000 construction examples, and roughly 35,000 floor plans contributing about two million drawing states. What comes out is a sequence of visual plans — a build order rendered as drawings. It is not an editable CAD feature history, and Skyfall lists feature-level output as future work.

That last distinction is the one an engineer will care about most. A sequence of pictures showing how a part is assembled is useful for review, for documentation, for training a downstream model, and for spotting an impossible order of operations. It is not a model you can open in a CAD package and edit, and it does not hand a machinist a toolpath.

Apollo World predicts the next second of traffic

Baidu's disclosure at AI Day·Apollo in Shenzhen was framed as a correction to the field. Wang Liang, chief architect of the Intelligent Driving Group, argued that a world model should not be a painter producing impressive video but an examiner verifying driving safety. The rubric that followed — "2+3" — is the clearest published attempt to define what counts.

The two gates are causality, meaning the next state is predicted from the current one rather than assembled frame by frame and back-filled, and interactivity, meaning an injected action changes what follows. The three graded properties are physical accuracy, persistence and real-time performance. Persistence is defined as the absence of drift in objects and details over long runs; real time is bounded by the one-thirtieth-of-a-second decision cycle, about 33 milliseconds, that a road vehicle needs.

The model's three capabilities map onto that. Traffic participants are agents with their own decisions rather than scripted props, so a braking robotaxi changes the behaviour of the cars behind it. Camera and lidar are generated in real time, so the stack under test perceives the virtual world through the same interface it uses on a real road. And trips close end to end — order, pickup, drop-off — across arbitrary origins and destinations. Baidu says the system already supports Apollo Go's global validation and cites around 100,000 simulated kilometres per day, framed as roughly three years of continuous testing by a safety-driver-equipped car. Apollo Go's operating figures — 28 cities, more than 23 million rides, 350 million autonomous kilometres, 240 million of them fully driverless — were quoted alongside.

No parameters, no architecture, no paper, no weights, no price. Apollo World is internal infrastructure that was described in public for the first time, not a product with a launch date you can buy against.

The numbers, and what a 43.6% IoU means

Skyfall published two benchmarks, both self-reported and neither independently replicated. On a CAD construction task of 1,000 components at roughly 100 drawing steps each, SKY-001 reports 3.9 seconds and $0.005 per case at a mean IoU of 79.7%. GPT-6 Astra is listed at 199 seconds, $0.75 and 79.1%; Gemini 4 at 253 seconds, $0.25 and 76.9%; Qwen 3.8 Max at 764 seconds, $0.30 and 34.4%; Kimi K3 at 439 seconds, $0.13 and 9.5%; Grok 4.7 produced no output in 1,250 seconds.

On a BIM floor-plan task of 1,000 plans at about 70 steps, SKY-001 reports 1.22 seconds, $0.0024 and 43.6% IoU, against Gemini 3.8 Flash at 42.3%, GPT-6 Astra at 40.6%, Kimi K3 at 33.7%, Qwen 3.8 Max at 28.8% and Grok 4.7 at 28.2%.

Read carefully, the tables say something more interesting than the headline. The CAD lead over GPT-6 Astra is 0.6 of an IoU point — Skyfall's own post says the model "matches or exceeds" rather than beats it — so the real gap is roughly 51 times faster and 150 times cheaper, not better output. The BIM lead is 3.0 points. And 43.6% mean IoU on floor plans means the best model in the table reconstructs less than half of the reference plan by intersection-over-union. That is a research-preview number, not a production number, and the landing page has been reported showing 76.1% where the blog shows 79.7% for CAD, which is the kind of drift you get when a figure is restated across surfaces.

One claim in the coverage does not survive contact with the sources: that SKY-001 beats Claude Opus 5.5. Opus 5.5 appears in the material as a blind language-model judge alongside GPT-6.1 Sol, not as a competitor on either benchmark. Being preferred by a judge in a qualitative comparison is not the same as scoring higher on IoU.

Where the comparison actually bites: who judges the output

The one methodological thread that connects the two announcements is evaluation by model. Skyfall's headline claims rest partly on a blind panel of language models scoring drawings. Baidu's claims rest on an internal rubric applied by its own engineers, with no external scoring at all. Neither company has published a protocol that a third party could rerun.

This is the part of the world-model boom that is least mature, and it is where a routing layer does real work rather than decorative work. Judging generated visual states is a model call — often several, from different vendors, so that no single lab's preferences define "good". OrcaRouter's model fusion lets a panel of models answer the same judging prompt together and returns their combined verdict through one endpoint, and the routing DSL can send different evaluation stages to different models without rewriting the harness. That is the layer we cover: more than 200 models, one key, 0% markup with provider list price passed through. Neither SKY-001 nor Apollo World is on that catalogue, and neither is a model you call — the judges are.

Neither is callable, and that changes the decision

The availability question has the same answer for both, for different reasons. SKY-001 is a research preview behind a waitlist, with commercial deployment targeted for early 2027 and no published terms. Apollo World is deployed inside Baidu and not offered externally. Neither has a price, and neither can be benchmarked against your own workload today.

What you can take from each is the method. SKY-001's two-stage decomposition — compose the whole, then predict the construction sequence of each leaf — is a reusable pattern for any process that has a hierarchy and an order of operations. Apollo World's 2+3 rubric is a reusable test for any simulator someone tries to sell you, and its persistence criterion in particular is the question most vendors cannot answer with a measurement.

Three questions that tell you which one you need

First: does the thing you are modelling have an agent acting inside it? If yes, you need an environment simulator and the Apollo World rubric applies. If the state transitions are deterministic and driven by a plan rather than a policy, you need something closer to SKY-001.

Second: is the output consumed by a human or by a control loop? A construction sequence is consumed by engineers and downstream models, so visual plausibility and step ordering are the metrics. A driving simulation is consumed by a perception and planning stack, so metric sensor consistency and latency are the metrics. The two have almost no evaluation criteria in common.

Third: do you need physics or do you need procedure? SKY-001 knows that a bracket must be attached before the assembly it holds; it does not know what happens when the bracket is dropped. Apollo World knows what happens when a pedestrian steps out; it has no opinion on how a part is manufactured.

The world-model label will keep stretching, because it is doing marketing work for at least four different product categories at once. The useful discipline is to ask what state is being predicted, what acts on it, and who consumes the prediction. Apollo World and SKY-001 answer those three questions in completely different ways, and once you have asked them the comparison dissolves into what it always was: two good ideas that happen to share a name.