A world generated one frame at a time
Odyssey has opened Odyssey-3 as a research preview, calling it the company’s “most powerful foundation world model yet”. It takes a text prompt and generates an interactive environment: the viewer navigates in first or third person, moves the camera independently, and can introduce events while the generation is still running.
The architecture is an autoregressive diffusion transformer — each new frame is predicted from the frames before it and from what the viewer just did. A distilled few-step variant is what makes the interaction run in real time rather than rendering offline.
Three tiers are named. The base Odyssey-3 outputs 832×480; Odyssey-3 Pro goes to 1280×720; Odyssey-3 Flash is the fastest. Odyssey has not published a price. Its own cost arithmetic assumes $1 per MI355X GPU-hour, excluding prompt-rewriting fees, and developers who want to build on the model are directed to a contact page rather than an API sign-up.
The benchmark numbers, and who produced them
Odyssey says Odyssey-3 Pro “sets a new state of the art on Physics-IQ Verified’s benchmark”, reporting 66.1 on the video-to-video task — which it describes as the highest reported score — and 54.7 on image-to-video. On WorldMark, averaged across 13 metrics, Odyssey says it ranks first in three of four categories, including first-person stylized at 79.0, ahead of HY-World 1.5 at 76.9, and third-person stylized at 76.3.
Two things qualify those figures, and both are visible in Odyssey’s own filings rather than dug out by a critic. The 66.1 came from a single run in which one video was selected out of eight generated. Without that selection, Odyssey-3 Pro averaged 63.37 across four runs. Odyssey submitted both results itself, in its pull request to the Physics-IQ benchmark repository. The WorldMark placements likewise come from Odyssey’s own evaluation, not an independent one.
That still leaves 63.37 as the clean number, and it is Odyssey’s to defend. But 66.1 and 63.37 are different claims, and only one of them describes what the model does when it is not allowed eight attempts.
What it is for
Odyssey frames physical accuracy as the point rather than visual polish — “We have made physical accuracy a central focus of our research,” the launch page says — which is why a physics benchmark is the one it leads with. A world model that predicts what happens next when something is pushed, dropped or driven into is useful well beyond entertainment: it is the same prediction a robot or a vehicle needs.
The preview on offer today does not do that. It covers interactive environment generation, and Odyssey says robotics and autonomous-systems applications still need further work.
What to watch
The thing to watch is an independent Physics-IQ run. The benchmark is maintained in a public Google DeepMind repository and submissions arrive as pull requests, so the disclosure is there for anyone who wants to check it — which is more than most vendor benchmark claims come with, even when the headline number flatters the model. Whether a third party reproduces 63.37, and what happens to the WorldMark placements under someone else’s evaluation, will say more than the launch page does.
