TL;DR
Odyssey opened a free public research preview of Odyssey-3, its “world model” that generates interactive 3D environments from a text description and simulates them in real time as you explore. The same model powers robot arms, a humanoid, drones, and a self-driving car — and the company claims top scores on physics benchmarks, though those numbers come with fine print worth knowing.
Type a world, then walk through it
The free online demo runs on Odyssey-3 Flash. You type a description, and the model builds an explorable environment around it — switchable between first-person and third-person views, with movement, camera control, and events you can trigger while the scene keeps generating. It is a research preview, not a product launch: the current build centers on environment generation, and the robotics and driving applications are demos of where the same model goes, not something you get access to yet.
Under the hood it is an autoregressive diffusion transformer: it predicts each new video frame from the frames before it and from your actions. Odyssey says it learns physical relationships and cause-and-effect from visual observation rather than from hand-built physics rules. The base model has 14 billion parameters and outputs 832-by-480 video; the higher-tier Odyssey-3 Pro runs at 1280-by-720. An extra training technique cuts the compute steps needed per frame enough to make the whole thing real time.
Training data came from three sources: internet videos with event descriptions, video game footage paired with the corresponding keyboard and mouse inputs, and simulated physical interactions. That mix is what connects what a scene looks like with how it changes when someone — or something — acts on it.
One model, many machines
The point of a world model is reuse: one pretrained brain, carried from one machine and task to another with only a small task-specific controller trained on top. As co-founder and CEO Oliver Cameron put it: “World models give physical agents a foundation of knowledge they can carry from one machine, environment, or task to another. We believe that the ability to learn broadly and then adapt with relatively little experience is the path toward increasingly general physical intelligence.”
The demos span six domains. In robotics, Odyssey-3 controlled multiple robot arms and recovered from failed grasps on its own, even though those recoveries were not in the training data. With Swiss company Flexion, controllers built on the model ran a humanoid in real time and handled lighting changes that broke the vision-language-action baselines Odyssey tested against. A car drove autonomously on real roads in India with a policy trained on roughly 20 hours of simulated driving data — no real-world driving footage. A drone navigated an indoor environment and dodged obstacles on simulated flight data. The model also played Grand Theft Auto V, and a controller trained on about two hours of GTA footage transferred its skills to Red Dead Redemption 2 with no extra training.
There is also a second-order play: Odyssey demoed an AI agent receiving natural-language tasks and learning inside an Odyssey-3-generated environment — a potential new way to train agents where they can fail and learn from consequences without touching the real world.
The benchmark claims come with fine print
Odyssey-3 Pro scored 66.1 on the Physics-IQ Verified video-to-video benchmark, which tests physical behavior across fluid mechanics, optics, solid mechanics, magnetism, and thermodynamics by having models continue videos of real-world experiments. But that record number came from a single test run where a selection method picked the best of eight generated videos per task — a procedure the benchmark's own rules do not accept for record claims, which require four runs with a reported standard deviation. Without the selection step, the model averaged 63.37 across four runs. Both numbers sit on the public leaderboard, but both were submitted by Odyssey itself.
On the WorldMark benchmark — control-following, image quality, world consistency — the company reports first place in three of four categories based on its own evaluation: 77.2 (first-person stylized), 79.0 (third-person real), and 76.3 (third-person stylized), with a third-place 80.6 in first-person real. Take these as company claims until independent evaluations land, but the interactive demo is open now — which means anyone can judge the most basic claim, that the worlds feel coherent, for themselves.
Why this matters
World models are the other bet on the road to general physical intelligence, parallel to the language-model ladder: instead of predicting the next word, the model predicts what the world looks like next and how it reacts to action. If one pretrained model really can be adapted to robots, cars, drones, and games with only hours of task-specific data, the cost of deploying capable physical agents drops sharply. That is why everyone is in this race: Google DeepMind is building Genie 3 along similar lines, and AMD agreed in late September to acquire Fei-Fei Li's World Labs for roughly $8.2 billion.
Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke, both from autonomous driving, and raised $310 million in June 2026 from investors including Amazon and AMD Ventures. The company has said developers can apply for API access to the model, so the next thing to watch is what third parties build on it — the demos so far were all run by Odyssey. For now, the free demo is the honest test: type a world, walk through it, and see whether it holds together.