Odyssey-3
Released Oct 8, 2026
Odyssey's foundation world model, an autoregressive diffusion transformer that predicts in real time how a text-prompted environment changes as a person or agent acts in it; also used to drive robot, vehicle, drone and game policies.
- Output
- Live video stream
- Interactive
- Yes
- Inputs
- text, actions
- Max session
- —
- Resolution
- 1280x720
- Frame rate
- —
- Access
- Research preview
- Licence
- Proprietary
Strengths & weaknesses
Strengths
- Real-time interaction via a distilled few-step variant
- Odyssey-3 Pro scores 66.1 on Physics-IQ Verified video-to-video (vendor-reported; claimed highest reported score)
- First- and third-person navigation plus agent-introduced events in the preview
Weaknesses
- Research preview only; no published developer pricing or API
- Low base resolution (832x480; 1280x720 only for Pro)
Announced 2026-09-15 ('Introducing Odyssey-3', said public release 'in the coming weeks'); research preview made available 2026-10-08 ('Our research preview is available now'), used as releaseDate. maxResolution is the Pro variant (base Odyssey-3 is 832x480, per a benchmark footnote). Physics-IQ and WorldMark figures are Odyssey's own evaluations. openWeights false = no weights mentioned anywhere. Predecessor Odyssey-2 Max (2026-04-21) was private beta for partners only. Odyssey also lists Agora-2 (multi-agent world model, 2026-09-24) with a demo at agora.odyssey.systems — not added as a separate entry. Notable=false only because access is 'research preview', not ga/beta, though the preview appears open to anyone — caller may wish to override. As of 2026-10-28 the Meet Odyssey-3 page shows 'Try Odyssey-3' and 'Get API Access' links. Odyssey's developer portal dates from the Odyssey-2 Pro API launch, and no Odyssey-3 model id or price was found.