Odyssey-3 turns a written description into an environment users can move through and affect in real time. Odyssey, a California-based AI company, has opened a public research preview of the model, giving people a way to explore its generated worlds while the company develops a wider set of uses.
From a prompt to an interactive environment
The free online demo runs on Odyssey-3 Flash. Users can describe an environment, then explore it from either a first-person or third-person perspective. They can move around, trigger events and see how the model responds.
That interaction is central to the idea of a world model. Rather than producing a fixed video, Odyssey-3 generates new frames in response to what has already appeared and to the user's actions. The company says the model simulates physical processes and predicts how an environment changes when someone does something within it.
Odyssey-3 is built on an autoregressive diffusion transformer. According to Odyssey, training teaches it physical relationships and cause and effect through visual observations. The training material included internet videos with event descriptions, game footage paired with keyboard and mouse inputs, and simulated physical interactions. An additional technique reduces the compute steps needed, which the company says makes real-time generation possible.
What the benchmark results show
Odyssey reports strong results on two benchmarks, but the figures need context. On Physics-IQ Verified, Odyssey-3 Pro scored 66.1 points in a video-to-video test covering areas such as fluid mechanics, optics, solid mechanics, magnetism and thermodynamics. The test asks models to continue videos of real experiments, with generated outcomes compared with what actually happened.
The 66.1-point result came from a single run using a selection method that chose one video from eight generations for each task. Benchmark rules call for four runs and a reported standard deviation to claim a record. Odyssey's result does not meet that requirement. Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs. Both results are on the official leaderboard, submitted by Odyssey.
On WorldMark, Odyssey says its own evaluation placed the model first in three categories: First-Person Stylized (77.2), Third-Person Real (79.0), and Third-Person Stylized (76.3). It placed third in First-Person Real with 80.6 points. WorldMark assesses instruction following, image quality and whether generated worlds remain consistent over time. These rankings provide one view of performance; the article reports that Odyssey submitted the leaderboard results itself.
A shared model for robots, drones and games
Odyssey's longer-term aim is to use one world model across different applications, pairing it with a specialized controller that turns predictions into commands. The examples range from robotic arms and drones to characters in video games.
In company tests, an AI system based on Odyssey-3 controlled multiple robotic arms after training on a few dozen hours of demonstrations. Odyssey says the arms could recover from failed grasps, including situations absent from the training data. For humanoid robots, the company is working with Swiss robotics company Flexion. Flexion-built controllers based on Odyssey-3 are reported to perform more reliably than comparison models when conditions change.
Odyssey also says it trained a drone controller using simulated flight data. In a virtual indoor setting, the drone avoided obstacles and flew to specified targets on command. In games, the company demonstrated an AI playing GTA V, steering vehicles and fighting enemies. A controller trained on roughly two hours of GTA footage also transferred its skills to Red Dead Redemption 2 without extra training, moving a character on horseback.
Another proposed use is training AI agents inside generated environments. In one demonstration, an agent received a task in natural language and attempted to complete it through its own actions. The goal is for agents to learn from the consequences of those actions. For now, the public research preview is focused on interactive environment generation; robotics and autonomous systems require further work.
Part of a broader push for world models
Founders Oliver Cameron and Jeff Hawke first unveiled Odyssey-3 on September 15, with robotics, autonomous driving and video games among its intended areas. Odyssey was founded in 2023. In June 2026, it raised $310 million from investors including Amazon and AMD Ventures.
Other organizations are pursuing similar technology. Google DeepMind is developing Genie 3 for generating interactive worlds, while World Labs, founded by Fei-Fei Li, is also working on world models. AMD announced in late September that it plans to acquire World Labs for roughly $8.2 billion.
The demo makes Odyssey-3's central proposition directly testable: a prompt can become a world that responds to movement and actions. The broader ambition is to make that kind of prediction useful for controlling systems. The gap between a compelling generated environment and reliable real-world robotics remains, by the company's own account, an area requiring further work.