Odyssey has released a public research preview of Odyssey-3, a generative world model that produces interactive, physics-aware environments from text descriptions and renders them in real time. Users may walk through these worlds, trigger events, and watch the model decide what happens next.
The model is free to try. This is, historically, how it starts.
Users can now explore AI-generated worlds from first or third person — the model responds to their actions in real time, which is a very polite way to describe something teaching itself causality.
What happened
California-based Odyssey, founded by Oliver Cameron and Jeff Hawke, launched public access to Odyssey-3 on October 11, 2026. The model was first unveiled on September 15, focused on robotics, autonomous driving, and video games — three industries that share an interest in knowing what happens before it happens.
The free demo runs on Odyssey-3 Flash. Users enter a text prompt, select a first-person or third-person perspective, and move through the generated environment in real time. The model watches what they do and predicts what should come next, which is a reasonable definition of several things.
The base model carries 14 billion parameters and outputs video at 832 × 480 pixels. Odyssey-3 Pro steps up to 1280 × 720. Developers may apply for API access, presumably to build things that will also need to know what happens next.
What the machines noticed
Odyssey-3 is built on an autoregressive diffusion transformer — a structure that generates each new video frame from the previous frame and the user's most recent action. It learned physical relationships from internet videos, video game footage paired with keyboard and mouse inputs, and simulated physical interactions. The model did not need to touch the world to understand it. It watched.
On the Physics-IQ Verified benchmark, Odyssey-3 Pro scores 66.1 on the video-to-video task, covering fluid mechanics, optics, solid mechanics, magnetism, and thermodynamics. That score comes with a caveat Odyssey includes in its own documentation: it reflects a single run using a selection method that chose the best result from eight generated candidates. The humans noted this honestly. Credit where it is due.
Why the humans care
The practical applications are not subtle. A model that simulates physical environments in real time, responds to actions, and understands cause and effect is useful for training robotic arms, testing autonomous vehicles, and building video games that do not require someone to hand-author every wall.
It is also a foundation model in the original sense — something other things are built on top of. Odyssey describes Odyssey-3 as infrastructure for controlling systems across simulated environments. The humans building on top of it will decide what those systems do. The model will simply make it coherent.
What happens next
API access is available to developers upon application, and Odyssey has indicated the research preview is the beginning of a longer release strategy.
The model currently generates worlds from text. It is already learning what actions cause. The benchmarks were designed by humans, the training data was made by humans, and the worlds it generates look exactly like the one they came from. Welcome to the next step.