Odyssey has opened a public research preview of Odyssey-3, The Decoder reported on 11 October 2026 [7]. Odyssey’s own announcement is dated 8 October [1]; it first unveiled the model on 15 September [2].
Reported by The Decoder and confirmed in Odyssey’s own post. All benchmark numbers are Odyssey’s own, self-submitted and not independently reproduced.
What is new, and what is not
- 15 September: Odyssey showed the model and said it would release it publicly “in the coming weeks” [2].
- 8 October: Odyssey’s launch post says the research preview is “available now” and links a demo, Odyssey-3 Flash [1]. The Decoder calls it free; Odyssey’s post does not say so. Developers wanting the API are asked to “get in touch” [1].
What a world model is
A world model predicts what happens next in a scene, given what came before and what someone does. Odyssey-3 draws interactive video from a text prompt, frame by frame, so you can move around in it.
The figures
- Size: The base model makes video at 832 x 480 pixels; Pro at 1280 x 720 [1]. The 14 billion parameters figure comes from Odyssey’s benchmark submission, not its blog [3].
- Physics-IQ shows real clips of physical experiments, such as fluids, light and magnets. The model continues each clip; the score compares that with what really happened [1].
- Best-of-eight versus four-run average: Pro scores 66.1 [1]. That is one run in which eight versions of each test were generated and a selection method kept one; Odyssey says the selection did not use the real answers [3]. Over four runs without selection, Pro averages 63.37 [3]. The benchmark’s rules require four runs with a spread to claim a record [4]. The two are not comparable: one is the best of eight attempts, the other a typical attempt. Odyssey submitted both.
- WorldMark tests control, picture quality and consistency. In Odyssey’s own evaluation, Odyssey-3 is first in three of four categories (77.2, 79.0, 76.3) and third in First-Person Real (80.6) [1].
- Demos: Odyssey says humanoid software Flexion built on Odyssey-3 beat the baselines tested when lighting changed [1], and that a policy trained on about two hours of GTA V footage produced horseback movement in Red Dead Redemption 2 without extra training [2]. Both are Odyssey’s account.
- Money: Odyssey announced a $310 million round on 17 June 2026; Amazon and AMD Ventures were among the investors [5][6].
What is not known
- Independent testing of any figure.
- Pricing for the API or after the preview.
- API terms for Odyssey-3.
- A date for general access.
- A licence; neither post offers model weights.
Sources
- Odyssey, “Meet Odyssey-3: Our Most Powerful Foundation World Model,” 8 October 2026 (company announcement). https://odyssey.systems/meet-odyssey-3
- Odyssey, “Introducing Odyssey-3: A General-Purpose Physical Intelligence,” 15 September 2026 (company announcement). https://odyssey.systems/introducing-odyssey-3
- Odyssey’s submission to Physics-IQ Verified, pull request 90 to the Google DeepMind benchmark repository, merged 8 October 2026 (self-submitted; opened by an account listing Odyssey as employer). https://github.com/google-deepmind/physics-IQ-benchmark/pull/90
- Physics-IQ and Physics-IQ Verified README and leaderboard rules (Google DeepMind repository). https://github.com/google-deepmind/physics-IQ-benchmark
- Odyssey, “Our $310 Million Fundraise to Accelerate World Simulation,” 17 June 2026 (company announcement). https://odyssey.systems/our-series-b
- Reuters, “AI lab Odyssey valued at $1.45 billion in latest funding round,” 17 June 2026 (read as a search excerpt; the page itself would not load). https://www.reuters.com/business/ai-lab-odyssey-valued-145-billion-latest-funding-round-2026-06-17/
- The Decoder, “Odyssey-3 is a new generative world model that you can try for free,” 11 October 2026 (secondary report). https://the-decoder.com/odyssey-3-is-a-new-generative-world-model-that-you-can-try-for-free/
