Odyssey opens a free preview of Odyssey-3’s interactive AI worlds

Odyssey has opened a free research preview of Odyssey-3, letting users describe an environment, move through it and change what happens in real time. The release makes its world-model approach publicly accessible, while the company’s larger ambition—using shared knowledge of physical behavior to control different machines—remains beyond what the public demo currently offers.

Founders Oliver Cameron and Jeff Hawke first showed Odyssey-3 on September 15, focusing on robotics, autonomous driving and video games. Odyssey followed with an October 8, 2026 announcement describing it as the company’s most powerful world model. The release now includes public access, technical details and benchmark results, moving the model beyond its initial introduction.

From a text prompt to an unfolding scene

The free online demo runs on Odyssey-3 Flash. Users can switch between first-person and third-person perspectives, navigate generated worlds and trigger events that the model responds to as they happen. Developers can apply for API access, a way for their software to interact with the model. The public preview currently centers on environment generation; robotics and autonomous-system applications require further work.

Underneath that interaction is an autoregressive diffusion transformer: a video-generating model that builds new frames from earlier frames and user actions. Odyssey says it learns physical relationships and cause and effect from visual observations. Its training combines internet videos with event descriptions, game footage paired with keyboard and mouse inputs, and simulated physical interactions. Those sources connect what a scene looks like with how it changes.

An additional training technique cuts the number of computing steps needed to generate the video, enabling real-time interaction. The base Odyssey-3 model has 14 billion parameters—the learned values inside a model—and generates video at 832 × 480 pixels, according to Odyssey’s benchmark submission. Odyssey-3 Pro supports 1280 × 720 pixels. The free demo, base model and Pro results therefore describe different members of the same model family.

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.

Oliver Cameron, Odyssey co-founder and CEO, in the October 8 announcement

The physics score needs its sampling conditions

Odyssey’s strongest disclosed physics result comes from Pro, not the Flash demo. Physics-IQ Verified asks models to continue videos of real experiments, then compares their generated outcomes with what actually happened. It covers fluid mechanics, optics, solid mechanics, magnetism and thermodynamics. The task tests whether a model’s continuation preserves physical behavior, rather than evaluating only whether the resulting images look convincing.

The company reported 66.1 points using a selection method that chose one of eight generated videos for each task. But that result came from a single run. The benchmark requires four runs and a reported standard deviation—the variation between results—for a record claim. The 66.1 result does not meet those requirements. Without that selection method, Pro averaged 63.37 across four runs. Both results are company submissions on the official leaderboard.

Odyssey also evaluated the model on WorldMark, which measures control-following, image quality and whether generated worlds remain consistent over time. In its own evaluation, Odyssey-3 ranked first in three of four categories: first-person stylized, third-person real and third-person stylized. It placed third in first-person real environments. That scorecard addresses interactive-world behavior separately from the experiment-continuation task used by Physics-IQ.

Controllers carry the model into other tasks

For machine and game applications, Odyssey pairs the world model with a specialized controller that turns its predictions into commands. The company says humanoid-autonomy business Flexion has adopted Odyssey-3. Its demonstrations also span smaller manipulation tasks, simulated flight and movement across games:

• Robot arms: Odyssey says a few dozen hours of demonstrations supported training across multiple arms, including recovery from failed grasps absent from training.
• Drones: A controller trained on simulated flight data avoided obstacles and flew toward commanded targets in a virtual indoor environment, according to the company.
• Games: Odyssey says a controller trained on roughly two hours of GTA V footage transferred to Red Dead Redemption 2 without additional training, moving a character on horseback.

Odyssey’s next proposed use is training AI agents inside generated environments. In one demonstration, an agent received a natural-language task and attempted it through its own actions in an Odyssey-3 scene. The goal is to let agents learn from the consequences of their actions. That remains a planned training use, distinct from the interactive environment generation available in the public preview.