ModelsArchitectures & capability
Odyssey opens free browser preview of Odyssey-3, a world model pitched for robots, cars and agent training
Odyssey-3 is a single world model that Odyssey says can be adapted to robot arms, humanoids, cars and agent training environments with tens of hours of data. A free browser preview has launched, and its top Physics-IQ score relies on best-of-8 selection.
QuietHow it’s measured
YouTube video · 1.2K views · 6 likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.
AT Pulse research deskPublished · 5 min read · AI-assisted analysisOdyssey describes Odyssey-3 as an autoregressive diffusion transformer that is pretrained once, then adapted to each machine with a small layer that turns its predictions into controls. Demos cover a robot arm, Flexion humanoids that kept working through lighting changes, and a car. On the car, the company reports that a driving policy built from 20 hours of data went about 77% as far between interventions as one trained on real footage. Sources disagree on whether those 20 hours were simulated or real road footage. [3] [8] [11] [14]
Physics-IQ Verified is a revised 2026 version of the Physics-IQ benchmark. On its video-to-video track, Odyssey-3 Pro scored 66.10% using best-of-8 sampling and custom prompts, and 63.37% from a single run. A different system ranks first on the image-to-video track. [4] [6] [13]
The browser research preview went live on October 8, 2026. Users can move through generated scenes and add events partway through. The base model runs at 832×480 and 16 fps, and the Pro model at 1280×720. API access requires a contact form, no weights have been released, and no official pricing has been published. [10] [11]
Odyssey-3 leads one Physics-IQ Verified track, scoring 63.37% from a single run.
Read the full assessment
Implication: if the partner robotics results hold up, teams could fine-tune one shared model instead of collecting large task-specific datasets, though no weights or public pricing exist yet.
Sources are checked. Writing the dossier now.
Executive brief
Odyssey's best Physics-IQ score is a best-of-8 result that the company submitted itself. Odyssey-3 Pro scored 66.10% on the video-to-video track of Physics-IQ Verified, but only after the model picked from eight samples using custom prompts. A single run scored 63.37% (DeepMind leaderboard). On the image-to-video track, other systems rank first. Odyssey-3 is one world model that the company says can run robot arms, humanoids, cars, drones and game agents, and a free browser preview opened on October 8, 2026. Most of the robotics results come from Odyssey or its partners and have not been checked independently.
What changed and event timeline
Physics-IQ benchmark introduced
Researchers from INSAIT and Google DeepMind released 396 real-world physics videos. They found that video models had very limited physical understanding, and that looking realistic did not mean being physically correct ().
Odyssey raises $310M
TNW reports the round valued Odyssey at $1.45B, with Amazon and AMD Ventures among the investors ().
Physics-IQ Verified published
Anates Labs, DeepMind and university co-authors fixed unclear prompts and artifacts in the ground-truth videos, and moved to per-sample scoring. Model rankings shifted a lot as a result (Kendall's τ=0.46) ().
Odyssey-3 unveiled
It was announced as a foundation world model for robots, humanoids, vehicles and drones, with public access promised "in the coming weeks" (;).
Research preview and leaderboard entries
The browser preview went live, and Odyssey submitted Physics-IQ Verified results through pull request #90 (;).
Creator explainer circulates
A Reddit post and an AI-avatar video summarized the launch alongside paid-community promotion ().
Capabilities and access
- Models: Odyssey-3 runs at 832×480 and 16 fps. Odyssey-3 Pro runs at 1280×720 (OrcaRouter).
- Preview: It runs in the browser at experience.odyssey.systems. Users can move in first or third person, and can add events partway through (Odyssey; video).
- API and weights: API access goes through a contact form. No weights or license have been released, and no official pricing has been published.
Read the full section
- Models: Odyssey-3 runs at 832×480 and 16 fps. Odyssey-3 Pro runs at 1280×720 (OrcaRouter).
- Preview: It runs in the browser at experience.odyssey.systems. Users can move in first or third person, and can add events partway through (Odyssey; video). The video says the preview uses a lighter, faster version of the model (05:05).
- API and weights: API access goes through a contact form. No weights or license have been released, and no official pricing has been published. OrcaRouter lists normalized per-video costs of $0.139 (base) and $0.267 (Pro) (OrcaRouter).
Technical analysis for researchers and developers
- Architecture: Odyssey describes the model as an autoregressive diffusion transformer. It has not published parameter counts, dataset sizes or a paper (Odyssey).
- Training data, per the video: three sources together: internet video, gameplay recordings paired with keyboard and mouse inputs, and simulated physics (04:41).
- How control works: a pretrained world model with a small machine-specific layer added on top that turns its predictions into controls (06:48).
Read the full section
- Architecture: Odyssey describes the model as an autoregressive diffusion transformer. It has not published parameter counts, dataset sizes or a paper (Odyssey).
- Training data, per the video: three sources together: internet video, gameplay recordings paired with keyboard and mouse inputs, and simulated physics (04:41). The gameplay data is what teaches the model how actions change a scene.
- How control works: a pretrained world model with a small machine-specific layer added on top that turns its predictions into controls (06:48).
- Benchmark method: the headline scores used best-of-8 sampling, chosen with WMReward and MBR consensus, plus custom prompts. Without selection, the scores are 63.37±0.63 (Pro) and 61.56±1.20 (base) (leaderboard). The Physics-IQ Verified paper does not say whether best-of-N is allowed (arXiv).
Claims and evidence
- Physics-IQ top score: 66.1 video-to-video and 54.7 image-to-video (Yahoo/press release). The 66.1 tops the video-to-video track on the leaderboard.
- Driving: policies trained on 20 hours of simulated data went about 77% as far between safety-driver interventions as policies trained on real footage (The AI Insider).
- Humanoids: Flexion's policies kept working through lighting changes that broke the VLA (vision-language-action) baselines it tested (Odyssey).
Read the full section
- Physics-IQ top score: 66.1 video-to-video and 54.7 image-to-video (Yahoo/press release). The 66.1 tops the video-to-video track on the leaderboard. On image-to-video, Physis-Lang with NVIDIA Cosmos 3 ranks first (Digg). OrcaRouter ranks the Pro model second by a different measure, points above the track mean (OrcaRouter).
- Driving: policies trained on 20 hours of simulated data went about 77% as far between safety-driver interventions as policies trained on real footage (The AI Insider). The video instead says the model trained on real Indian road footage (02:58). CEO Oliver Cameron's post says only "20 hours of driving data" (X).
- Humanoids: Flexion's policies kept working through lighting changes that broke the VLA (vision-language-action) baselines it tested (Odyssey). This comes from a partner, and the baselines are not named.
- TNW: reports first place in three of four WorldMark categories, an evaluation Odyssey ran itself (TNW).
Context and prior work
The original Physics-IQ study found that Sora, Runway, Pika, Lumiere and others had very limited physical understanding (INSAIT). Odyssey's team comes from self-driving: Cameron says he built self-driving cars from 2015 to 2023 (X; video 01:12). Competitors include NVIDIA's open Cosmos models and World Labs, which AMD acquired for a reported $8.2B (TNW).
Limitations, safety and contested findings
- No other group has reproduced the robotics, driving or drone results. They come from Odyssey's own materials and partners.
- The robot arm's recovery from a missed grab is shown as a demo, with no success rates attached (02:09).
- The video's claim that Odyssey-3 is "the most physically accurate world model" (05:44) goes further than the leaderboard supports.
Read the full section
- No other group has reproduced the robotics, driving or drone results. They come from Odyssey's own materials and partners.
- The robot arm's recovery from a missed grab is shown as a demo, with no success rates attached (02:09). The video's narrator says they have not tested it themselves (06:25).
- The video's claim that Odyssey-3 is "the most physically accurate world model" (05:44) goes further than the leaderboard supports. That holds for one track, with best-of-8 selection.
- The early three-camera result, multi-view output after 100 training steps, appears only in the video (05:13).
- The captions are automatic. They garble "Anates Labs" and drift into unrelated promotional content near the end.
Business and practitioner implications
- Robotics and autonomy teams: the pitch is fine-tuning one shared model with tens of hours of data instead of collecting large task-specific datasets.
- Agent builders: generated worlds could serve as cheap training environments. Check how well behavior learned there carries over to the real world.
- Procurement: with no weights, no published pricing and only form-based API access, there is lock-in risk.
Read the full section
- Robotics and autonomy teams: the pitch is fine-tuning one shared model with tens of hours of data instead of collecting large task-specific datasets. Test that on your own hardware before planning around it.
- Agent builders: generated worlds could serve as cheap training environments. Check how well behavior learned there carries over to the real world.
- Procurement: with no weights, no published pricing and only form-based API access, there is lock-in risk. Benchmark the single-sample scores, since best-of-8 multiplies compute cost.
The source trail.
Sources (15)
AI World Models Explained: Why Odyssey-3 Matters
Feed excerpt only. A transcript of the video linked from this page is supplied (linked video auto captions; https://www.youtube.com/watch?v=zSfcDHJMeVo); automatic text may contain errors.
reddit.comOdyssey-3 Can Train AIs AND Control Robots
Related coverage; assess separately
reddit.comLinked video: AI World Models Explained: Why Odyssey-3 Matters
linked video auto captions
www.youtube.com