Apollo World vs Odyssey 3: Two Ways to Teach a Car Inside a Dream

Odyssey-3 is the only model in this comparison with a price on it, so start there: roughly $0.139 for a generated video at 832 by 480, or $0.267 at 1280 by 720 on the Pro tier, released as a free research preview on 8 October 2026 and available through a contact form rather than a self-serve key. Apollo World has no price, because Baidu has not put one on it. The company disclosed the model on 3 November 2026 as an agent-driven simulator that trains and grades its own robotaxi fleet, and published no endpoint, no weights, no parameter count and no commercial terms. Both are world models. Both have been demonstrated training a driving policy. Only one of them is a product, and it is not the one built by the company that operates the cars.

That inversion is the story. Odyssey is a venture-backed startup selling imagination by the clip; Baidu is an operator using imagination to avoid putting a safety driver in a real car. They are solving the same problem from opposite ends of the market.

What $0.139 actually buys

Odyssey-3 is an autoregressive diffusion transformer of roughly 14 billion parameters, trained on internet video, gameplay footage with time-aligned input, and simulated rigid-body interactions. It is distilled into a few-step model so that a person can navigate the output in real time: you prompt an environment, then move through it in first or third person, with independent camera control and the ability to trigger events.

• Base tier — 832 by 480, about $0.139 per generated video, 16 frames per second in the benchmark harness.

• Pro tier — 1280 by 720, about $0.267 per video.

• Flash — the variant behind the free browser demo, which anyone can try without an account.

• Access — a free public research preview in the browser; API access by contact form; no self-service keys, no published per-token rate card, no open weights and no licence.

The lack of self-service is a real constraint, not a footnote. You cannot benchmark Odyssey-3 against your own workload this afternoon the way you can with an open-weight model. You can try the demo, and you can talk to sales. For a team evaluating world models as infrastructure, that is a different kind of friction from a price.

The 20-hour Indian road experiment

The demonstration that matters for this comparison is a driving policy trained on about 20 hours of real Indian road data with a frozen backbone, inside Odyssey's generated worlds. The company also showed robot-arm manipulation learned from tens of hours of demonstrations, including recovery behaviours that were not in the training data, humanoid policies built with Flexion, drone navigation, and a GTA V-trained policy transferring to Red Dead Redemption 2.

Those demos are the strongest argument for a general video world model as a training substrate: if a policy can learn from footage of a game and transfer to another game, the model has captured something about dynamics rather than about pixels. But they are demonstrations, not measurements. There is no published success rate on the Indian road task, no baseline against training in the real world, and no ablation showing which part of the generated world did the work. The claim is that it is possible. The claim is not that it is better.

Baidu's approach is the industrial version of the same idea, with the evaluation turned inward. Apollo World models every traffic participant as an agent that decides rather than a prop that follows a script, so a robotaxi braking or yielding provokes a reaction from the cars and pedestrians around it. It generates camera and lidar in real time, so the stack under test perceives the virtual city through the same interface it uses on a road. And it closes trips end to end — order, pickup, drop-off — across arbitrary origins and destinations.

The scale claim is around 100,000 kilometres of simulated driving per day, which Baidu frames as roughly three years of continuous testing by a safety-driver-equipped car. The operating context is Apollo Go: 28 cities, more than 23 million rides, 350 million autonomous kilometres and more than 240 million fully driverless ones. Every one of those numbers is Baidu's own, and none has been independently audited.

The benchmark arithmetic worth learning from

Odyssey published a headline number for Odyssey-3 Pro on Physics-IQ Verified, a video-to-video physical reasoning benchmark: 66.1, described as state of the art. The caveat is in the method. That score came from a single run using best-of-eight selection, while the benchmark's rules call for four runs with a reported standard deviation. Without selection, the four-run average was 63.37.

Both numbers are Odyssey's. The gap between them — 2.73 points — is the cost of picking the best of eight samples and calling it the score. On the benchmark's own measure of net improvement over the track average, Odyssey-3 Pro ranks second at +11.15 percentage points, behind FLUX 3 [large] at +12.27. On WorldMark, Odyssey's own evaluation, the model is first in three of four categories — 77.2 first-person stylised, 79.0 third-person real, 76.3 third-person stylised — and third in first-person real at 80.6, behind Lyra 2.0 at 84.4 and AlayaWorld at 83.0.

None of this is a scandal. Publishing a best-of-eight number and the four-run average, as Odyssey's coverage did, is more transparency than most labs offer. But it is the reason a comparison between Odyssey-3 and Apollo World on benchmark scores is not possible: one side has numbers with a documented method and a dispute about selection, and the other has no numbers at all. The honest table would have a column of values and a column of dashes, and a table with a column of dashes is not a comparison.

Two definitions of real time

Both models claim real-time operation, and the two claims are not the same quantity. Odyssey-3 runs at 16 frames per second in its benchmark harness at 832 by 480 — fast enough for a person to steer a camera through the world without waiting, which is what the product needs. Apollo World is built around a decision cycle of one thirtieth of a second, about 33 milliseconds, because a road vehicle travelling at speed cannot wait longer than that to act.

Sixteen frames per second is 62.5 milliseconds per frame. That is not a criticism of Odyssey, which is not claiming to be a vehicle simulator; it is an illustration of how differently the two systems are specified. A model that a human explores can afford to be a little late. A model that closes a control loop with a two-tonne car cannot.

The same split shows up in persistence. Odyssey's own evaluation puts it third in first-person real-world scenes, and long-horizon drift is the classic failure of autoregressive video: the world slowly stops matching itself. Baidu lists persistence as one of three graded properties and defines it as the absence of drift in objects and details over long runs, which is a specification written by someone who has watched a simulation degrade over ten hours.

Who can use which, and what it costs

Odyssey-3 is usable today by anyone with a browser, and by anyone with a sales conversation. At $0.139 to $0.267 per video, the unit economics are legible for content work and for research that needs a few thousand clips. It is a commercial product with a free tier, and the absence of open weights is a deliberate choice rather than an oversight.

Apollo World is usable by nobody outside Baidu today. There is no indication in the 3 November material that this changes, and the disclosure was framed entirely around Apollo Go's own testing and validation.

Pricing comparisons across the two are therefore not like-for-like in any dimension. Odyssey bills per generated video; Baidu bills nobody. And for the models we do carry, OrcaRouter's rule is the plain one — 0% markup, provider list price passed through, so a vendor price change is live the same day. Odyssey-3 is not in that catalogue, so there is no pass-through rate to quote, and this article is not going to invent one.

The verdict, such as it is

If you want to train a policy in imagination this quarter and you do not work at Baidu, Odyssey-3 is the more accessible instrument, with the caveat that "accessible" means a demo and a contact form. Its driving-policy demonstration is the most directly relevant public evidence that a general video world model can serve as a training substrate for embodied control, and the company's willingness to publish both the headline and the four-run average is a point in its favour.

If you are specifying a simulator rather than buying one, Apollo World's 2+3 rubric is the more valuable artifact: causality, interactivity, physical accuracy, persistence and latency under complexity. Ask Odyssey, or any vendor, how their model scores on persistence over a ten-hour run, and whether the physics benchmark number was selected or averaged. The answers will tell you more than any leaderboard position.

What would settle the comparison is a third party running both on the same driving task with the same policy and the same metric. Nobody has, and neither company has proposed it. Until then, the accurate summary is a startup selling dream footage by the clip and an operator using dreams to keep its cars off the road — similar ideas, opposite business models, and no shared scoreboard.