Odyssey has opened the interactive demo of its physical intelligence model Odyssey-3 to everyone

In the race of AI models that build worlds from text, an unusual step was taken this week. Odyssey, which has reached a $1.45 billion valuation, opened an interactive demo of its Odyssey-3 model, which it describes as “general-purpose physical intelligence,” to everyone through the browser. A user types a text; the model generates a real-time environment that can be explored in first-person, third-person, and free-camera modes, and is open to adding events and actions. The demo, following the company’s September 15 blog announcement and October 8 launch, takes the model out of report pages and turns it into a usable tool.

One backbone, six different embodiments

Odyssey positions its new architecture not as a video model but as a world model. According to the company, a single pre-trained backbone carries knowledge of physics, dynamics, causality, and human behavior to different embodiments; the model can adapt to a new task with little experience. The areas listed at launch are humanoid robot and robot arm control, vehicle driving, drone piloting, synthetic environment generation for autonomous agents, and video games. Flexion, which works on humanoid robot autonomy, is among early users.

Founder Oliver Cameron also builds the claim on approaching a single intelligence that can understand and manage its environment and learn from broad data and adapt with little experience.

The benchmark table and the limits of the numbers

The measurements point in the same direction. In the Physics-IQ Verified test, the model scored 66.1 in the single best run, the highest individual result seen in this table; the average of four runs is 63.37. Sources remind that the difference between the best run and the average should be read carefully. It was reported as first place in three categories of the WorldMark evaluation, and third place with 80.6 in the first-person reality category. The information that the model has 14 billion parameters, however, comes from a single source and was not confirmed by the company.

The access gap with Genie 3

At the other end of the comparison is Google DeepMind’s Genie 3. Announced on August 5, 2025, the model produces 720p resolution and 24 fps; interaction is limited to a few minutes, and access in the US is primarily behind a Google AI Ultra subscription. Odyssey’s demo presentation, on the other hand, is free and opens in the browser by application; however, there is no API, and session limits and pricing were not disclosed. The Flash variant produces at 832×480, and the Pro variant at 1280×720.

Financial picture

The financial map also confirms where Odyssey stands. In mid-June, a $310 million Series B round led by Natural Capital took the company to a $1.45 billion valuation; Amazon, GV, AMD Ventures, EQT, and In-Q-Tel joined the round, along with angel investors Jeff Dean, Kyle Vogt, and Garry Tan. Total funding reached $337 million.

In the end, world models are coming down from the research lab to the product shelf. Competition is no longer measured by on-paper frame rates and benchmark scores, but by how many seconds it takes for a single sentence typed by a user to produce an explorable world. The demo is free; it is not yet clear what robotics, driving, and simulation customers will pay for their own training environments.

Source: WebTuring