Odyssey opens its Odyssey-3 world model to free public testing
Odyssey opened a free public research preview of Odyssey-3, a world model that generates interactive environments from text prompts in real time. The demo runs the Flash variant, while the 66.1 physics score Odyssey cites came from its Pro model on a best-of-eight run.
The Engineer · Build desk
Two bars comparing Odyssey-3 Pro's benchmark scores: 66.1 points from a single run where a selection method picked one of eight generated videos per task, and 63.37 points averaged across four runs without the selection method.
Odyssey-3 Pro, same benchmark, results Odyssey submitted to the leaderboard In points
| Item | Value | Claim |
|---|---|---|
| One of eight picked, single run | 66.1 points | 10 |
| Four-run average, no selection | 63.37 points | 11 |
What happened
- Founders Oliver Cameron and Jeff Hawke first showed Odyssey-3 on September 15, aiming it at robotics, autonomous driving and video games.
- Odyssey-3 is an autoregressive diffusion transformer that keeps generating new video frames from the previous frames and the user's actions.
- Odyssey's own WorldMark evaluation puts Odyssey-3 first in three of four categories, with scores of 77.2, 79.0 and 76.3 in them.
- Swiss robotics company Flexion built humanoid controllers on Odyssey-3 that are said to be more reliable than comparison models when conditions change.
Why it matters
- constraint Public hands-on testing reaches the Flash variant only, so Pro's physics standing still rests on leaderboard entries Odyssey submitted itself.
- decision Teams comparing Odyssey-3 Pro with rival models should take the four-run, no-selection average as the comparable figure, since an interactive session produces one stream of frames with nothing to select from.
- cost Putting Odyssey-3 on a robot means building a task-specific controller and collecting demonstration data to train it; Odyssey's arm result used a few dozen hours.
Odyssey credits the real-time speed to an extra training technique that cuts the number of compute steps the model needs [8]. We think the training mix is the best-judged part of the release. Alongside internet video with event descriptions and simulated physical interactions, Odyssey used game footage paired with the keyboard and mouse inputs that produced it [7]. A model that predicts each frame from the user's action [6] needs footage with the action recorded beside it, and game capture provides exactly that.
The specs in the account are for the base model and Pro. According to the official benchmark submission, the base model has 14 billion parameters and renders 832x480 [9]. Pro supports 1280x720 [9], about 2.3 times as many pixels per frame [14]. The account does not give Flash's size or resolution, or say which variants developers who apply for API access will get [4].
Physics-IQ Verified has models continue videos of real-world experiments in fluid mechanics, optics, solid mechanics, magnetism and thermodynamics, then compares the output with the real outcome [17]. Pro's 66.1, as reported by Odyssey [16], came from a single run [10]. The benchmark's rules require four runs with a standard deviation for any record claim, and The Decoder reported that the entry does not meet that bar [10]. Without selection, Pro averaged 63.37 over four runs [11]. That is 2.73 points lower [13]. For the higher figure to transfer, a team's workload would have to generate eight candidates per task and include a selector that picks well [15]. That is a lot to ask of a robot arm mid-grasp. Both results are on the official leaderboard, and Odyssey submitted both [11].
WorldMark checks how well a model follows control instructions, how good its images look and whether its world stays consistent over time [19]. In our view consistency is the hardest of the three for an autoregressive model, because its later frames are built on frames it generated itself [6]. The one category Odyssey-3 does not lead, First-Person Real, holds its highest score, 80.6, and there it ranks third [18].
Odyssey also says arms controlled through Odyssey-3 recovered from failed grasps on their own, though such failures were not in the training data [21]. Recovery from an unseen failure is the robotics result we would most want reproduced. An environment demo cannot test it [5].
What to watch
- Whether Odyssey files four runs with a standard deviation for its selection setup, the condition Physics-IQ Verified sets for a record claim.
- Which Odyssey-3 variants the application-only API serves, and Flash's parameter count and output resolution.
- Independent WorldMark evaluations of Odyssey-3, or Flexion publishing the metric behind its reliability comparison.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence40
- Adoption15
- Hype gap+35
- Incentives70
- Confidence45
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
California-based AI company Odyssey has launched a public research preview of its world model Odyssey-3.
- [2]
Odyssey-3 generates interactive environments from text prompts that users can explore in real time.
- [3]
Founders Oliver Cameron and Jeff Hawke first unveiled the model on September 15, with a focus on robotics, autonomous driving and video games.
- [4]
Developers can apply for API access to Odyssey-3.
- [5]
The free online demo runs on Odyssey-3 Flash; users can choose first- or third-person perspective, move through the generated world, trigger events and watch the model respond in real time.
- [6]
Odyssey-3 is built on an autoregressive diffusion transformer that continuously generates new video frames based on previous frames and user actions.
- [7]
Training data included internet videos with event descriptions, video game footage paired with the corresponding keyboard and mouse inputs, and simulated physical interactions.
- [8]
An extra training technique reduces the number of required compute steps, making real-time generation possible.
- [9]
The base Odyssey-3 model has 14 billion parameters and generates video at 832x480 pixels, according to the official benchmark submission; Odyssey-3 Pro supports 1280x720 pixels.
- [10]
The 66.1 score is from a single test run where a selection method picked one of eight generated videos for each task; the benchmark rules require four test runs with standard deviation reported for any record claim, and the reported record does not meet that bar.
- [11]
Without the selection method, Odyssey-3 Pro averaged 63.37 points across four runs; both results appear on the official leaderboard but were submitted by Odyssey itself.
- [12]
For each application, the world model is paired with a specialized controller that translates its predictions into concrete commands, across robotic arms, drones and game characters.
- [13]
Odyssey-3 Pro's best-of-eight single-run Physics-IQ Verified score exceeds its four-run no-selection average by 2.73 points.
- [14]
Odyssey-3 Pro's 1280x720 frames carry about 2.3 times the pixels of the base model's 832x480 frames.
- [15]
Reproducing the best-of-eight score means generating eight videos per task, eight times a single-sample run, plus a selection method to choose among them.
- [16]
According to Odyssey, Odyssey-3 Pro scores 66.1 points on the video-to-video benchmark from Physics-IQ Verified.
- [17]
Physics-IQ Verified evaluates physical behavior across fluid mechanics, optics, solid mechanics, magnetism and thermodynamics; models continue videos of real-world experiments and outputs are compared against actual outcomes.
- [18]
On WorldMark, based on its own evaluation, Odyssey-3 ranks first in First-Person Stylized (77.2), Third-Person Real (79.0) and Third-Person Stylized (76.3), and third in First-Person Real with 80.6 points.
- [19]
WorldMark tests how well models follow control instructions, how good the generated images look, and whether simulated worlds stay consistent over time.
- [20]
According to Odyssey, an AI built on Odyssey-3 controlled multiple robotic arms, and a few dozen hours of demonstration data was enough for training.
- [21]
According to Odyssey, the robots could recover from failed grasps on their own, even though those situations were not part of the training data.
- [22]
Odyssey is working with Swiss robotics company Flexion on humanoid robots; controllers Flexion built on Odyssey-3 are said to perform more reliably than comparison models, even when conditions change.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
- the-decoder.comOdyssey-3 is a new generative world model that you can try for free
1 article · October 11, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.