What is Odyssey-3?
Odyssey-3 is the latest foundation world model from Odyssey, an AI lab founded by veterans of the self-driving industry. Unveiled on September 15, 2026, it is an autoregressive diffusion transformer trained on a vast collection of visual observations of the world, from which it has learned physics, dynamics, cause-and-effect and human behaviour.
At heart it is still a generative world model: it predicts how the world unfolds, frame by frame, and can simulate interactive environments where people and AI agents act and see the consequences. What is new is how widely Odyssey wants to use that knowledge.
One model, many machines
Instead of training a separate system for every robot or vehicle, Odyssey trains small control policies on top of the same pretrained world model. According to the company, Odyssey-3 learned simulated driving from roughly 20 hours of data, and humanoid skills from just a few dozen hours of demonstrations.
In a collaboration with humanoid specialist Flexion, policies built on Odyssey-3 opened containers, moved and placed objects, and generalised better than the vision-language-action baselines they were compared against — continuing to work under lighting changes that made those baselines fail.
From Odyssey-2 to Odyssey-3
Odyssey made its name with Odyssey-2 in October 2025, a causal model that streams “interactive video” frame by frame in response to your inputs, which the team compared to a GPT-2 moment for world models. It was followed by Odyssey-2 Pro, with a public developer API, and Odyssey-2 Max. Odyssey has also raised a $310 million round, with AWS as a preferred cloud partner. Odyssey-3 extends that foundation from interactive video to physical intelligence.
Availability
Odyssey-3 has not been released yet; the company plans public availability in the coming weeks. No speed, latency, resolution or hardware figures have been published, and all results so far come from Odyssey's own testing. Odyssey-2 Pro remains available today through the developer API.
Strengths and limitations
Strengths
- One pretrained model for robots, vehicles, drones and games
- Learns new physical tasks from little task-specific data
- Focus on cause-and-effect, a weak spot of most world models
- Proven interactive-video lineage with a public API (Odyssey-2 Pro)
Limitations
- Not released yet
- No published speed, resolution or hardware figures
- Results so far are company-reported
Who is Odyssey-3 for?
Robotics and autonomy teams looking for a shared world-model foundation, and developers following the move from interactive video toward physical AI.
Frequently asked questions
Is Odyssey-3 available?
Not yet. Odyssey announced it on September 15, 2026 and plans a public release in the coming weeks. Odyssey-2 Pro is available today through the developer API.
Is Odyssey-3 a world model or a robot controller?
It is a generative world model. Robot, vehicle and game controllers are trained as policies on top of it, using its learned understanding of the world.
How much data does Odyssey-3 need to learn a new task?
Odyssey reports about 20 hours of simulated driving data for driving, and a few dozen hours of demonstrations for humanoid skills.
Is Odyssey-3 real-time?
Odyssey has not published speed or latency figures. Its predecessor, Odyssey-2, streams interactive video in real time.