The model is built on an autoregressive diffusion transformer, which continuously generates new video frames based on previous frames and user actions. Odyssey says it learned physical relationships and cause and effect from visual observations, using training data that included internet videos with event descriptions, video game footage paired with keyboard and mouse inputs, and simulated physical interactions. An extra training technique cuts the number of compute steps needed, making real-time generation possible.

The base model has 14 billion parameters and generates video at 832 × 480 pixels, with the more powerful Odyssey-3 Pro supporting 1280 × 720. On benchmarks, the picture needs a caveat: Odyssey-3 Pro's claimed top score of 66.1 on the Physics-IQ Verified video benchmark came from a single test run in which a selection method picked one of eight generated videos per task. The benchmark rules require four runs with standard deviation reported, so the record claim does not meet that bar. Without the selection method, the model averaged 63.37 across four runs. Both figures appear on the official leaderboard, but were submitted by Odyssey itself.

Beyond virtual worlds, Odyssey positions Odyssey-3 as a foundation for controlling systems, pairing the model with specialised controllers for tasks such as robotic arms, drones and game characters. The company says an AI built on the model controlled multiple robotic arms after training on a few dozen hours of demonstration data, and could recover from failed grasps it had never seen in training. Odyssey also showed the model autonomously playing GTA V, steering vehicles and fighting enemies.

Founders Oliver Cameron and Jeff Hawke first unveiled the model on September 15 with a focus on robotics, autonomous driving and video games. What is new now is public access, more technical detail and the benchmark results.