has launched a public research preview of its world model Odyssey-3. The model generates interactive environments from text prompts in real time, letting users explore and interact with them from different perspectives.
A world model is an AI system that learns to predict how a scene changes, including how actions affect what happens next. World models learn from large amounts of visual data to predict how a scene will change next, including how actions affect what happens. Odyssey also says one model can steer machines, not just generate worlds.
What the preview offers
The free online demo runs on Odyssey-3 Flash, a faster variant. Users can pick first-person or third-person views, move around and trigger events. The weights are closed, and API access is by request.
According to the benchmark submission cited by The Decoder, the base model has 14 billion parameters, the internal settings a model learns in training, and it generates video at 832 × 480 pixels. The Pro version supports 1280 × 720. The preview runs the Flash variant, not the 720p Pro model behind the headline score.
Odyssey describes Odyssey-3 as "our most powerful foundation world model yet" in its launch post. Secondary reports date the launch to October 8. Odyssey first showed the model on September 15.
The physics score and its caveat
Odyssey says Odyssey-3 Pro scored 66.1 on the video-to-video test of Physics-IQ Verified. The test asks models to continue videos of real physical experiments and compares their predictions with what actually happened. The company calls this the highest reported score.
The Decoder says the figure needs context. It comes from a single run in which a selection method picked the best of eight generated videos per task. The benchmark's rules require four runs with standard deviation reported for any record claim, so the result does not meet that bar. Without the selection step, Pro averaged 63.37 across four runs. Odyssey submitted both results to the leaderboard itself.
Other coverage describes the same setup. One report says the comparison model FLUX 3 [large] scored 64.35, also on a top-eight basis. A separate summary of Odyssey's footnote says the scores average four runs, which differs from The Decoder's account of the 66.1 figure.
On the WorldMark benchmark, by Odyssey's own evaluation, the model ranks first in three of four categories. In first-person real scenes it places third, behind Lyra 2.0 and AlayaWorld.
Robots, drones, cars and games
Odyssey pairs the world model with a specialised controller for each task. All the results below are the company's own claims.
Robot arms: tens of hours of demonstrations were enough to train them. The company says they showed recovery behaviors that were absent from the demonstrations. Humanoids: Swiss robotics company Flexion built control policies on Odyssey-3. Odyssey says they beat the tested baselines when the environment changed. Driving: Odyssey says it trained a driving policy for real roads in India on 20 hours of driving data, with the underlying model left unchanged. Drones: a controller trained on simulated flight data dodged obstacles in a virtual indoor space. Games: a controller trained on about two hours of GTA V footage played the game. It also moved a horseback character in Red Dead Redemption 2 with no extra training.
The Decoder notes that the public preview covers interactive environment generation for now. Robotics and autonomous systems need further work.
The wider race
Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke. The Decoder reports that it raised $310 million in June 2026 from investors including Amazon and AMD Ventures. Google DeepMind is pursuing a similar approach with Genie 3, and World Labs, founded by Fei-Fei Li, is developing comparable technology. AMD said in late September that it plans to acquire World Labs for roughly $8.2 billion.
Odyssey is also pitching the model as a place to train AI agents. In one demo, an agent received a task in natural language and tried to complete it inside a generated world.
