The Pulse
Odyssey Launches Odyssey-3 for Interactive World Simulation
Odyssey released Odyssey-3 on October 8, 2026, with a research preview for developers building physical AI systems. The company says its Pro version scored 66.1 on Physics-IQ Verified’s video-to-video benchmark, while cautioning that benchm
AI.info Team · Written with AI ·
Strong benchmark scores, with a boundary
Odyssey says its new world model leads on selected simulation benchmarks, but the company also cautions that those scores do not show how well the system will perform on a particular robot or task. The distinction matters: Odyssey-3 is presented both as a tool for generating interactive environments and as a foundation that developers can adapt to physical systems.
Released on October 8, Odyssey-3 is an autoregressive diffusion transformer that predicts how objects move and interact and how situations change over time. Odyssey says the model learns representations of physics, dynamics and cause and effect from visual observations, then uses them to simulate environments or help train policies for machines.
“These results measure specific properties of generated worlds; applying the model to a physical system also requires evaluating the behaviors that matter for that machine and its tasks,” write Oliver Cameron and Jeff Hawke of Odyssey.
Odyssey-3 Pro scores 66.1 on Physics-IQ
Odyssey reports that Odyssey-3 Pro scored 66.1 on Physics-IQ Verified’s video-to-video benchmark, the highest reported score on the leaderboard cited in the announcement. The benchmark tests predictions of real experiments in fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. Odyssey-3 Pro also scored 54.7 on the benchmark’s image-to-video task.
Those figures are specific to the benchmark and model variant: Odyssey’s page identifies the 66.1 result as Odyssey-3 Pro’s video-to-video score, not a general measure of physical capability. The company says standard Odyssey-3 improves the measured tradeoff between physical accuracy and generation cost, allowing more simulations within the same compute budget.
Odyssey also evaluated the model on WorldMark, which measures control-following, visual quality and world memory. Using the mean of the benchmark’s 13 reported metrics, the company says Odyssey-3 ranked first in three of four categories: first-person stylized, third-person real and third-person stylized environments. It placed third in first-person real.
A prompt becomes a world that responds
The research preview lets users generate an environment from a prompt and move through it while the model predicts how it changes in response to actions or new events. Odyssey says users can switch between first- and third-person navigation and move the camera independently. That makes the release more than a system that produces a fixed video: its stated purpose is to continue a scene as people or agents act within it.
The model’s training data combines internet video with time-localized event annotations, gameplay recordings paired with keyboard and mouse inputs, and simulated rigid-body interactions with captions and metadata. Odyssey describes training it to continue from prior observations and predict future states given prompts and controls, then applying a post-training pipeline to create a faster variant for real-time interaction.
From generated scenes to machines
Odyssey’s examples of physical-system adaptation depend on additional components trained for each machine. For robot arms, the company says tens of hours of demonstrations supported manipulation tasks and recovery actions, including reorienting a gripper after a missed grasp and retrieving a dropped object.
For driving, Odyssey says it kept the model’s backbone frozen while training a policy on 20 hours of driving data to operate a car on real roads in India. The policy uses visual representations from Odyssey-3 to predict waypoints. The company also describes work with humanoid policies developed by Flexion, but the benchmark results alone do not establish how those systems will perform beyond the tested conditions.
The research preview is available now. Odyssey is asking developers working on robots, humanoids, self-driving cars, drones and other autonomous systems to contact the company about building with Odyssey-3.