Odyssey 3

Odyssey-3 is trying to be more than another model that spits out text or images. It’s a foundation world model that builds interactive environments from a prompt, then predicts how those environments change in real time as a person or agent acts inside them.

That alone makes it interesting. But the bigger hook is the Pro version, which says it has the highest reported Physics-IQ Verified video-to-video score: 66.1, best-of-8. The number matters because this is not just about pretty scenes. It’s about whether the model can keep up when the world starts moving.

The same model has also been adapted for very different jobs. According to the launch, it can control robot arms and humanoids, drive a car, and train agents. That’s a wide spread, and it suggests the team is pitching Odyssey-3 as a general platform for action, not a narrow demo.

There’s a research preview available now, and API access is being offered by request. So this is both a public showing and a quiet invitation to builders who want to push on the model rather than just watch it perform.