Overview
Odyssey is an AI lab building foundation world models — generative systems that predict the next frame of a scene given your current view and whatever action you just took, rather than returning one fixed clip for one prompt. The distinction matters: where a normal video model hands everyone the same output, Odyssey returns a system you act inside. You press a key, tap a screen or move a joystick, and the model streams what it thinks happens next, at up to 30 FPS with around 40 ms of latency in its real-time preview. The company positions this as an early holodeck — interactive entertainment, film previsualization, robotics training and simulation all sit in the same capability.
Odyssey-3, released on 7 October 2026, is the most powerful iteration to date. On the Physics-IQ Verified leaderboard it ranks near the top of a physical-plausibility board previously owned by cinematic generators, and it does so at roughly a third of the normalized cost of competitors like FLUX 3. The Pro tier (720p) posts the highest image-to-video score on that board. Independent analysts note the honest caveat: the entry is a self-submission, and in first-person real environments Odyssey-3 places third, behind Lyra 2.0 and AlayaWorld. That is the right kind of evidence for the claim — a frozen backbone plus a small adapter yields meaningful behavior from tens of hours of robot or driving data — but it is vendor-reported, not independently replicated.
For our evaluation the interesting part is not ‘best video model in the world’ but ‘a world model built for interaction has landed near the top of a physics board.’ The action decoder attached to the frozen model produced recovery behaviors never in the demonstrations, which is exactly what makes the physical-system results worth taking seriously.
Key Features
- Interactive world model — generates the next frame from your current state plus an action, so you steer the scene live instead of re-prompting
- Real-time streaming — up to 30 FPS from GPU clusters in the US and EU, with ~40 ms latency between input and returned frame
- First- and third-person camera modes — accept an event mid-generation and keep world memory across the shot
- Physics-IQ leadership — Odyssey-3 and Pro top the verified physical-plausibility board at a fraction of rival normalized cost
- Robot and driving transfer — frozen backbone plus small adapters reportedly generalize to manipulation, humanoid and closed-loop driving tasks
- Open research posture — transparent self-submitted benchmark rows and public preview access for builders to test
Pricing
| Plan | Price | For |
|---|---|---|
| Research Preview | Free (waitlist) | Anyone wanting to try interactive video in-browser |
| API / credits | TBD | Developers; infrastructure runs ~$1–2 per user-hour today |
| Enterprise / licensing | Custom | Studios, game and sim teams needing integration |
Odyssey has no self-serve pricing page yet — access is via the research preview waitlist, and production API pricing is still to be announced. Treat any per-video cost as infrastructure-level, not a committed product rate.
Comparison
Compared to Kling, Odyssey is not a prompt-to-clip generator — Kling excels at polished short films you render once, while Odyssey gives you a world you drive frame by frame. Against Runway, the same split holds: Runway is the editing and effects studio, Odyssey is the simulation you inhabit. Next to Sora, Odyssey trades cinematic breadth for physical control and interactivity, which is why it tops physics benchmarks rather than general creative leaderboards.