The launch
On October 11, 2026, California startup Odyssey opened a public research preview of Odyssey-3, a world model that generates interactive environments in real time from text. First shown on September 15, what's new is free public access, more technical detail and benchmark results. Developers can apply for API access.
How it works
- The free demo runs on Odyssey-3 Flash. Users describe a world, pick first- or third-person, move around, trigger events and watch the model respond.
- Architecture: an autoregressive diffusion transformer generating each frame from previous frames and user actions.
- Trained on internet video with event descriptions, game footage paired with keyboard and mouse inputs, and simulated physical interactions.
- The base model has 14 billion parameters at 832×480; Odyssey-3 Pro supports 1280×720.
Benchmarks, with an asterisk
Odyssey says Pro scored 66.1 on Physics-IQ Verified (video-to-video), which tests fluids, optics, mechanics and thermodynamics. But as The Decoder showed, that record comes from a single run picking the best of eight videos per task; the rules require four runs with standard deviation. Without selection, the average was 63.37. WorldMark results (1st in three of four categories) are also self-evaluated.
One model for robots, drones and GTA V
The plan is one world model with specialized controllers:
- Robot arms: a few dozen hours of demos sufficed, and robots recovered from failed grasps on their own (per the company).
- Humanoids: partnership with Switzerland's Flexion.
- Drones: a controller trained on simulated flight dodged obstacles in a virtual environment.
- Games: an AI played GTA V, and with about two hours of footage transferred skills to Red Dead Redemption 2 without extra training.
- Agents: the goal is training agents that learn from consequences of their actions inside generated worlds.
Robotics and autonomy aren't part of the public preview yet.
Context
Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke and raised $310 million in June, with Amazon and AMD Ventures among investors. The world-model race includes Google DeepMind's Genie 3 and Fei-Fei Li's World Labs, which The Decoder says AMD plans to acquire for about $8.2 billion.
Why it matters for businesses
- Cheap simulation for robotics: world models could cut real-world data collection for warehouse and factory robots.
- Skepticism toward self-reported benchmarks: check the average and method, not just the record.
- Start with the demo: it's free and shows real limits (resolution, consistency over time).
- Still research: for production, plan months of physical validation.
Sources
- The Decoder (2026-10-11): https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/
Want to evaluate world models and simulation for robotics or agent training? KGA IT offers a free consultation: kga-it.com.