World model Odyssey-3 can now be tested for free in the browser, with API applications for developers
The world model from the California company Odyssey opened a public research preview on October 8, 2026, with real-time generation from text prompts and an API application process for developers. At the same time, it turns out that the top result on Physics-IQ Verified comes from a single run with selection — not the four runs with standard deviation that the benchmark's rules require for record claims.
The concrete news
The California AI company Odyssey has launched a public research preview of its world model Odyssey-3. Users can now generate interactive environments from text prompts for free and explore them in real time in the browser. The model's demo runs on Odyssey-3 Flash, and developers can apply for API access (The Decoder).
What is new, then, is access. The founders Oliver Cameron and Jeff Hawke unveiled the model back on September 15, with a focus on robotics, autonomous driving and video games. What came in addition with the October 8, 2026 launch is public access, more technical detail and the first benchmark results.
How the model works
Odyssey-3 is built on an autoregressive diffusion transformer that continuously generates new video frames based on previous frames and the user's actions. According to Odyssey, the model learns physical relationships and cause and effect from visual observations during training, with training data from internet video with event descriptions, gameplay video paired with inputs, and simulated physical interactions (The Decoder).
The base model has 14 billion parameters and generates video at 832 × 480 pixels, according to the official benchmark submission. Odyssey-3 Pro supports 1280 × 720 pixels (The Decoder).
The record result that breaks the rules
This is where the picture becomes more nuanced than the launch headlines suggest. On the Physics-IQ Verified benchmark, developed by Anates Labs and DeepMind, which asks models to continue videos of real physics experiments, Odyssey-3 Pro delivers a top result of 66.1 points — with models from Nvidia and Black Forest Labs in the comparison (The Next Web).
But The Decoder's review of the leaderboard data shows that the record result stems from a single test run in which a selection method picked one of eight generated videos per task. The benchmark's rules require four test runs with reported standard deviation for any record claim, and the 66.1 result does not meet that requirement. Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs. Both results appear on the official leaderboard, but both were submitted by Odyssey itself (The Decoder).
On the WorldMark benchmark, Odyssey-3 ranks first in three of four categories — based on the company's own evaluation: First-Person Stylized (77.2), Third-Person Real (79.0) and Third-Person Stylized (76.3). In First-Person Real, the model places third with 80.6 points (The Decoder).
The company's claims — and what is verified
In its press release via Business Wire, a paid distribution channel, Odyssey presents Odyssey-3 as the company's most powerful world model to date, and claims the model shows shared world understanding across robots, humanoids, vehicles, drones, AI agents and video games (Odyssey via Business Wire).
Among the concrete claims, all of which come from Odyssey's own press release and have not been independently verified:
- The model is said to have driven autonomously on roads in India with a driving policy trained on roughly 20 hours of simulated driving data.
- It is said to control multiple robotic arms and to have shown recovery behavior it was not explicitly trained on.
- Odyssey runs control policies developed together with the humanoid company Flexion.
- The model is said to play Grand Theft Auto V, with early transfer of behavior to other games without additional policy training.
Odyssey's co-founder and CEO Oliver Cameron describes the model this way in the press release (the quote is translated from English): "Odyssey-3 is a big step toward a single intelligence that can understand and operate in the world around us" (Odyssey via Business Wire).
A necessary caveat
An independent review in Yasuhito Morimoto's Physical AI News newsletter points out that each machine type requires dedicated control policies, and that a single model does not directly operate all the machines without additional design (Morimoto). According to Morimoto, the drone demonstration stems from the September unveiling and took place in simulation, not as a new physical result from the October launch.
Money and market
Odyssey raised $310 million in June at a valuation of $1.45 billion, with Amazon and AMD Ventures among the investors. Competitors include World Labs, Emulate and Nvidia's open Cosmos platform (The Next Web).
What remains unresolved
Two questions remain after the launch: whether Odyssey will publish rule-compliant four-run results on Physics-IQ Verified that can confirm the record claim, and whether the performance claims — the India driving, the robotic arm demonstrations, the GTA V transfer — can be independently verified. For now, the safest move is to try the demo yourself and read the benchmark numbers with the methodological caveat they require.