Introduction

Odyssey-3 is the new foundation world model from Odyssey, the world-model lab led by Oliver Cameron and Jeff Hawke. It launched on 2026-10-08: type a prompt and it generates an interactive environment in real time, then keeps predicting how that world changes as you walk through it or drop in an event. A free research preview (served by a fast variant Odyssey calls Odyssey-3 Flash) is live in the browser at experience.odyssey.systems.

Modality: prompt to interactive video world (game-like, playable, real time). Open vs API: closed weights; the browser preview is free, and API access is by request through Odyssey's developer site. The

passed 250k views in its first few hours.

What you can do in the preview

Prompt a scene, then act in it. Today's preview gives you three ways to move:

  • First-person navigation
  • Third-person navigation, with a character in frame
  • Free camera movement, so you can inspect what the model predicted from a different angle

You can also introduce an event mid-generation and watch how the world responds. Each new frame is predicted from what came before plus your latest input, so this is a continuous simulation rather than a set of separate video clips.

Under the hood

Odyssey-3 is an autoregressive diffusion transformer. Odyssey trained it as a multi-step video diffusion model with prompts and controls that change over time, extended it autoregressively (teacher forcing plus causal masking) so it continues from earlier frames conditioned on actions, then distilled it into a few-step model fast enough for real-time play.

The training mix is interesting for game makers. Alongside internet video with time-stamped event annotations and simulated rigid-body physics, it includes gameplay recordings paired with keyboard and mouse inputs, which is why it understands controls at all.

The numbers

Benchmark (Odyssey's reported results)Odyssey-3 (832×480)Odyssey-3 Pro (1280×720)

Physics-IQ Verified, video-to-video

51.8 base, 64.4 best-of-8

63.4 prompt-enhanced, 66.1 best-of-8

Physics-IQ Verified, image-to-video

41.0 base, 52.8 best-of-8

54.7 best-of-8

Odyssey says the 66.1 is the highest reported score on the Physics-IQ Verified leaderboard as of Oct 7, ahead of the FLUX 3 and NVIDIA Cosmos3 entries charted beside it. On WorldMark (control-following, visual quality and world memory), Odyssey's own evaluation puts Odyssey-3 first in three of four splits: first-person stylized (77.2, a hair over LingBot-World's 77.0), third-person real (79.0) and third-person stylized (76.3). It comes third on first-person real (80.6) behind Lyra 2.0 (84.4) and AlayaWorld (83.0). Treat the WorldMark ranks as vendor-run until the benchmark publishes them.

The game angle

Two things matter beyond the playable demo.

Agents that play real games. Odyssey trained game-playing policies on top of the frozen world model using gameplay recordings plus keyboard and mouse inputs. It reports long play sessions in GTA V and early transfer: a movement policy trained on about two hours of GTA footage produced horseback riding in Red Dead Redemption 2, and GTA policies rode motorcycles in Sleeping Dogs with no extra training on either game.

Worlds for training agents. The same model generates environments where an AI receives a goal in plain language and works toward it. That's the idea behind Odyssey's PROWL-2 research, and its separate Agora-2 model already lets several humans or AIs share one live simulated world.

Caveats

This is a research preview, not a game engine. There's no exported scene, mesh or save file, only a generated stream you steer, and the launch posts don't state a session length limit or how long the world remembers what you've seen. The preview runs the faster Flash variant, not the 720p Pro model behind the headline score. API access runs through Odyssey's developer site and a get-in-touch flow pitched at robotics and physical-AI teams; the launch pages list no price, and there are no open weights. If you need weights to run locally, the open-weights option we covered this week is WorldPlay2.

Original Source

Conclusion

Odyssey-3 is the easiest way right now to feel where real-time world models are heading: open a tab, type a scene and walk around inside it. For game developers the bigger signal is the training recipe. A model fed gameplay paired with controls ended up driving agents in GTA V and carrying that skill into Red Dead Redemption 2, which hints at future tools for automated playtesting and believable NPCs. Try the free preview this weekend, and keep an eye on whether the API opens up to game studios.

—Titus