Odyssey released its Odyssey-3 world model on September 15, 2026, with a research preview becoming available on October 8. The model is an autoregressive diffusion transformer designed to predict physical interactions and object movements based on visual data. It utilizes three data sources: labeled internet videos, gameplay recordings, and physics simulations.
The company offers two versions: Odyssey-3 with 832×480 resolution and Odyssey-3 Pro with 1280×720 resolution. The Pro version achieved a score of 66.1 on the Physics-IQ Verified benchmark using a best-of-8 approach. Odyssey also evaluated the model on the WorldMark benchmark, which measures control adherence, visual quality, and world memory.
Odyssey-3 serves as a foundation for training control policies for various systems, including robotic manipulators, humanoids, and autonomous vehicles. For example, policies trained on 20 hours of simulated driving were tested in real-world street environments. The company is also developing projects like PROWL for recursive agent training and Agora-2 for multi-agent environments.
The developers acknowledge that world models currently lag behind large language models in scale by approximately two orders of magnitude. While the model demonstrates capabilities in simulated environments, the company notes that real-world performance remains a primary challenge. Public pricing, licensing, and API documentation have not been disclosed.
