Odyssey-3 opens to the public: Real-time interactive worlds from text

According to a report dated 11 October 2026, California-based Odyssey has opened access to its Odyssey-3 world model as a public research preview. The system generates real-time interactive environments from text prompts, allowing users to move through them from different perspectives and trigger events. The free online demo runs on the Odyssey-3 Flash model, while developers must apply for API access.

Exploring worlds described in text

Odyssey-3 is built on an autoregressive diffusion transformer that generates new video frames based on previous frames and user actions. The company says the model learns physical relationships and cause-and-effect connections from visual observations. Its training used internet videos paired with event descriptions, game footage combined with keyboard and mouse input, and simulated physical interactions.

Demo users can choose a first- or third-person perspective, move through the generated environment, and see how the system responds to their actions. According to the company, an additional training technique that reduces the number of computation steps makes real-time generation possible.

In its official benchmark post, the company describes the base Odyssey-3 model as a 14-billion-parameter system that generates video at 832 × 480 pixels. Odyssey-3 Pro is listed at 1280 × 720 pixels.

An important caveat about benchmark results

Odyssey reports that Odyssey-3 Pro scored 66.1 on the Physics-IQ Verified video-to-video test. However, this figure is based on a single test run in which one of eight videos generated for each task was selected. The benchmark rules cited in the source require four test runs, with the standard deviation also reported, to claim a record. The average across four runs without the selection method is 63.37.

In the WorldMark evaluation, Odyssey’s own test results put the model in first place in three of four categories: 77.2 in First-Person Stylized, 79.0 in Third-Person Real, and 76.3 in Third-Person Stylized. It ranks third in First-Person Real, with a score of 80.6. The test examines areas such as instruction following, the quality of generated images, and consistency of the environment over time. These results should be understood as based on the company’s own evaluation.

Targets in robotics and game control

Odyssey plans to use its world model for a range of purposes, from directing robotic arms and drones to controlling game characters. The model’s predictions are converted into commands through application-specific controllers. The company says a system based on Odyssey-3 controls multiple robotic arms and can recover from failed grasps not shown during training. Work on humanoid robots is being carried out in partnership with Swiss robotics company Flexion.

A drone controller trained by the company on simulated flight data navigates obstacles and moves toward targets in a virtual indoor environment. Another example shown by Odyssey features AI driving a vehicle and fighting enemies in GTA V. The company also says a controller trained on around two hours of GTA footage was able to move a character on horseback in Red Dead Redemption 2 without additional training.

Implications for architecture and visualization workflows

The open preview currently focuses on generating interactive environments. The source provides no information about integration with architectural design or visualization software, so there is no basis for treating Odyssey-3 as a replacement for existing production tools. Still, the ability to move through text-described spaces and see how actions affect an environment may be worth exploring for teams developing ideas for interactive scenes.

For studios and 3D artists, testing should include output resolution, how consistently the environment responds to actions, and compatibility with their software. API access requires an application; pricing and access terms have not been disclosed. The source also does not specify hardware requirements. For robotics and autonomous systems, the preview should not be taken as a ready-to-deploy solution; the company says work in these areas is ongoing.