Odyssey-3 opens a free public preview of its world model
The world-model race added a third serious entrant you can actually touch this weekend — Odyssey opened its Odyssey-3 to the public, and the benchmark claims came with strings attached.
Odyssey-3 is now a public research preview, and the free demo runs today. California-based Odyssey — the 2023 lab founded by Oliver Cameron and Jeff Hawke — first showed off Odyssey-3 in mid-September; what's new is public access, a published technical write-up, and numbers. The demo at Odyssey's experience site runs on the smaller Odyssey-3 Flash: you type a description, pick first- or third-person, and the model generates the environment in real time, letting you walk through it and trigger events. Under the hood it's an autoregressive diffusion transformer that keeps producing new frames from the last frames plus your inputs, trained on internet video with event captions, game footage paired with keyboard and mouse inputs, and simulated physics. The base model is 14 billion parameters at 832 × 480; Pro goes to 1280 × 720. Developers can apply for API access.
The benchmark record is real — but it doesn't meet the benchmark's own bar. Odyssey-3 Pro posted 66.1 points on Physics-IQ Verified, which would top the leaderboard, and the company submitted the result itself. The catch: that score came from one run where a selection step picked the best of eight generated videos per task, while the benchmark requires four runs with a reported standard deviation for any record claim. Without the selection trick, Odyssey-3 Pro averaged 63.37 across four runs — still strong, just not a clean record. On the WorldMark control benchmark, Odyssey claims first place in three of four categories on its own evaluation. Our take: shipping the caveats alongside the score is more honesty than we usually see in world-model marketing, and the gap between 66.1 and 63.37 is exactly why leaderboard-watching needs the fine print.
The bigger bet: one model for robots, drones, and game characters. Odyssey's pitch is that the same frozen backbone, paired with a task-specific controller, can drive robotic arms (trained on a few dozen hours of demos, recovering from failed grasps it never saw), humanoid robots with Swiss firm Flexion, indoor drones, and even game characters — a controller trained on about two hours of GTA V footage reportedly transferred to Red Dead Redemption 2 with no extra training. The public preview covers interactive generation only; the robotics side is still, in Odyssey's own words, further work. Odyssey raised $310 million from Amazon, AMD Ventures and others in June, and it's chasing the same prize as Google's Genie 3 and Fei-Fei Li's World Labs — which AMD agreed to buy for roughly $8.2 billion last month. If you need a primer on what separates a world model from a video generator, we explained the distinction in AI 101 — What is a world model?.
What to watch: whether the Physics-IQ leaderboard accepts the submission as filed, and whether the API access push brings actual developer adoption.
Is a self-reported benchmark win with disclosed caveats worth celebrating, or should the leaderboard reject best-of-eight runs outright? Tell us in the comments.
