Odyssey-3 world model opens a research preview, posting the top reported Physics-IQ score and real-time interactive environments

Why it matters

【Key points】

World-model startup Odyssey launched Odyssey-3 on Oct 8. The model was first announced in mid-September; this release makes it usable: a research preview (Odyssey-3 Flash) is open to try now, and physical-AI developers can request API access.

Odyssey-3 is an autoregressive diffusion transformer that learns how objects move and situations evolve from broad visual data. From a prompt it generates explorable environments in real time and predicts how they change as a user moves or introduces events, with first-person, third-person and free-camera modes.

Odyssey reports that Odyssey-3 Pro scores 66.1 on Physics-IQ Verified video-to-video, the highest reported score (720p, best-of-8), and 54.7 image-to-video. In its own WorldMark evaluation it ranks first in 3 of 4 categories. Showcased uses: robot-arm manipulation from tens of hours of demos, including recovery behaviors not in the demos; Flexion humanoid policies that held up under lighting changes where tested VLA baselines failed; and driving on real roads in India with a policy trained on 20 hours of data.

【Verifiable】

Odyssey blog: https://odyssey.systems/meet-odyssey-3

Official X: https://x.com/odysseyml/status/2108238378751967385

Physics-IQ Verified leaderboard: physics-iq-verified.anates.ai (Oct 7 snapshot cited by Odyssey)

【Our read】

1. The real change is from announced to usable: a world model outsiders can try, placed on a third-party physics leaderboard.

2. The 66.1 top score uses the Pro model with best-of-8 and prompt enhancement, and the WorldMark ranking is self-run. Read the numbers with their settings, not as single-shot quality.

3. The robot, humanoid and driving results are still company demos without independent replication. The thing to watch is whether outside teams, once API access opens, can train useful policies from little data.