Blog
Odyssey opens a free public preview of its Odyssey-3 world model
Odyssey has opened a free browser research preview of Odyssey-3, a world model that generates explorable environments from text in real time. Its headline physics benchmark score comes with caveats, and the robotics uses it has shown are not in the public preview.
California startup Odyssey has opened a free public research preview of Odyssey-3, a "foundation world model" that generates an environment from a text prompt and lets you move through it in real time. The preview launched around 8 October 2026; Odyssey first unveiled the model on 15 September.
What you can try
The browser preview runs on the lighter Odyssey-3 Flash model. According to Odyssey and early coverage:
- You type a prompt and the model generates an interactive scene.
- You can switch between first-person and third-person navigation and move the camera independently.
- You can trigger events and watch the model respond.
Hands-on testers report a queue and short timed sessions of roughly five minutes, with no login required. One review found scenes lost structural consistency when the camera turned around, a common weakness for this type of model.
API access is not self-serve. Developers have to apply through Odyssey's contact form, and no public pricing or open weights have been announced.
How it works
Odyssey describes Odyssey-3 as an autoregressive diffusion transformer: it keeps generating new video frames based on earlier frames and the user's actions. According to The Decoder, training data included internet video with event descriptions, game footage paired with keyboard and mouse inputs, and simulated physical interactions. Odyssey-3 Pro outputs 1280×720.
The benchmark caveat
Odyssey says Odyssey-3 Pro set a new state of the art on the Physics-IQ Verified video-to-video benchmark with a score of 66.1. As The Decoder reports, that figure came from a single run that picked the best of eight generated videos per task. The benchmark's rules require four runs with a reported standard deviation for record claims. Run that way, Odyssey-3 Pro averaged 63.37. Both results were self-submitted by Odyssey.
Odyssey also says, based on its own evaluation, that Odyssey-3 ranks first in three of four WorldMark categories and third in the fourth (First-Person Real).
Robots, drones and games: not in the preview
Odyssey's bigger pitch is that one world model, paired with task-specific controllers, can drive many systems. It has shown:
- Robotic arms trained on a few dozen hours of demonstration data, which it says recovered from failed grasps
- Humanoid controllers built with Swiss company Flexion
- A drone controller trained on simulated flights
- An agent playing GTA V that transferred to Red Dead Redemption 2 without extra training
These are company demonstrations. None of them is part of the public preview, and Odyssey says robotics and autonomous-systems use needs further work.
Why it matters
World models are a growing bet across the industry. Google DeepMind has Genie 3, and The Decoder notes AMD announced plans in late September to acquire World Labs. Odyssey raised $310 million in June 2026 from investors including Amazon and AMD Ventures. A free preview lets anyone judge the technology directly instead of relying on demo videos, which is useful given the benchmark asterisk.