Odyssey-3 Generates Virtual Worlds in Real Time — and You Can Walk Through Them Right in Your Browser

Odyssey has opened public access to Odyssey-3, a new world model that generates an environment from a text description and keeps rebuilding it as you interact with it. Unlike a typical video generator, it lets users move through the world they created, change the camera position and add new events while generation is still running. The research version of Odyssey-3 Flash runs in a browser with no local installation. odyssey.systems

The public launch took place on October 8. As Odyssey explains, the model takes previous frames and new user actions into account, trying to predict how the scene should change next. First-person and third-person navigation are available, along with separate camera controls. odyssey.systems

That is a fundamental difference from a pre-generated clip: the image is not simply played back but keeps being created as the user explores the environment.

Odyssey-3 Tops a Physics Comprehension Test

Odyssey-3 is built on an autoregressive diffusion transformer. To train it, the company used internet video annotated with events, gameplay recordings with synchronized keystrokes and mouse movements, and simulations of rigid-body interactions. The model’s job is not only to produce a convincing image but also to account for object motion, cause-and-effect relationships and previous events. odyssey.systems

The most notable result came from Odyssey-3 Pro on Physics-IQ Verified, a benchmark that tests the physical behavior of generative video models. It uses recordings of real experiments with fluids, rigid bodies, light, magnetic phenomena and other physical processes.

Odyssey-3 Pro took first place in the video-to-video category with a score of 66.10. There is an important caveat, though: the record figure was achieved using the best-of-8 method, which picks the best of eight generations. In a standard run with the same user prompts, Pro scores about 63.37. The base Odyssey-3 scored 61.56 in a similar configuration. These values have already been added to the official Physics-IQ Verified leaderboard.

In Odyssey’s own testing, the model also took first place in three of the four sections of WorldMark, which evaluates command following, visual quality and world-information retention. At the same time, the developers themselves stress that such benchmarks test only individual properties of generation and do not prove a universal understanding of real physics. odyssey.systems

Today we’re launching Odyssey-3, the most powerful foundation world model yet. It sets a new state of the art on Physics-IQ, and powers robots, trains AIs, and generates interactive experiences. It’s really cool. Experience the model today, all for free! https://t.co/LHWa7MPbKi

— Odyssey (@odysseyml) October 8, 2026

From Virtual Worlds to Robots and Cars

For Odyssey, interactive worlds are just one use of the model. The company is already adapting it to control physical devices.

In experiments, Odyssey-3 was used for robotic manipulators, and Odyssey’s partner Flexion built a humanoid robot control system on the model. For the automotive scenario, the developers trained a control policy on 20 hours of data, leaving the base Odyssey-3 unchanged, after which the car was tested in a closed loop on real roads in India. odyssey.systems

The developers see such world models as a virtual environment where robots and AI agents can gain experience without having to run costly physical tests every time. But the published demos remain research experiments rather than proof that the system is ready for mass use.

Regular users get access to Odyssey-3 Flash, while the Physics-IQ record belongs to the heavier Pro version at 1280 × 720 pixels. So the capabilities of the public browser preview should not be compared directly with the results of the flagship model.

Odyssey-3 also cannot yet be considered a replacement for Unity, Unreal Engine or another full-fledged game engine: the model generates a changing video world, not a finished 3D scene with familiar objects, geometry and game systems. But the ability to create an environment and explore it right away shows where generative AI may go after images and ordinary video.