Odyssey launched Odyssey-3 on 8 October 2026, calling it “our most powerful foundation world model yet”, and opened a research preview. The company says Odyssey-3 Pro sets a new state of the art on Physics-IQ Verified’s video-to-video benchmark with a score of 66.1, “the highest reported score.”
Its launch post on X added: “Experience the model today, all for free!” The model was first introduced in an Odyssey post on 15 September 2026.
Odyssey describes Odyssey-3 as “a learned dynamical system, implemented as an autoregressive diffusion transformer, that predicts how objects move and interact through space and how situations evolve over time.”
The research preview lets a user prompt an environment and move through it. Odyssey says it “provides first-person and third-person navigation alongside independent camera movement”, and that a user can introduce an event during generation and watch how the model responds.
Odyssey invites physical-AI developers to “get in touch” to build with the model. The posts we read give no price for developer access.
WorldMark, mean of 13 metrics per split, as reported in Odyssey’s own evaluation: Odyssey-3 leads three of four splits and is third in first-person real, behind Lyra 2.0 and AlayaWorld.
Business Pill · PHYSICAL AI
A one-minute explainer of physical AI: models that perceive and act in the physical world through robots, vehicles and other machines. It teaches the general idea only and says nothing about any company in this story.
The key insight: As we read it, Odyssey is pitching Odyssey-3 less as a video generator than as a shared starting point for machines: one pretrained model, then tens of hours of each machine’s own data to adapt it, with the benchmark scores and demonstrations so far coming from Odyssey’s own runs.
The Benchmark Claims
On Physics-IQ, which Odyssey describes as a benchmark from Anates Labs and DeepMind, Odyssey says Odyssey-3 Pro scores 66.1 in video-to-video and 54.7 in image-to-video. The test asks models to continue videos of real physical experiments across fluid dynamics, optics, solid mechanics, magnetism and thermodynamics.
Odyssey’s footnote says scores average 4 runs, cites the Physics-IQ Verified leaderboard of 7 October 2026, and gives the resolutions: 832×480 for Odyssey-3 and 1280×720 for Pro. For costs, it says Odyssey “assumes $1 per MI355X GPU-hour, excluding prompt-rewriting fees.”
On WorldMark, which measures control-following, visual quality and world memory, Odyssey reports its own evaluation, using the benchmark’s captions and the mean of 13 metrics. Odyssey-3 ranks first in three of four splits: 77.2 in first-person stylized, 79.0 in third-person real and 76.3 in third-person stylized. In first-person real it scores 80.6, third behind Lyra 2.0 at 84.4 and AlayaWorld at 83.0.
Odyssey adds its own caveat: “These results measure specific properties of generated worlds; applying the model to a physical system also requires evaluating the behaviors that matter for that machine and its tasks.”
One Model, Several Machines
Odyssey says the model’s knowledge is adapted to each machine by training an action decoder or policy on paired observations and actions. It reports these demonstrations:
Robot arms: “With only tens of hours of robot demonstrations, Odyssey-3 completed manipulation tasks and showed recovery behaviors absent from those demonstrations.”
Humanoids: Odyssey says Flexion built humanoid control policies on Odyssey-3, and that they “exceeded the performance of the tested VLA baselines under environmental changes.”
Driving: Odyssey says it trained a driving policy for real roads in India “on just 20 hours of driving data while keeping the Odyssey-3 backbone frozen.”
Sensors: in what it calls an early experiment, an Odyssey-3 training checkpoint produced three-camera driving sequences “after just 100 training steps.” The 15 September post also names drone piloting and a collaboration with Poke & Wiggle to evaluate the model across different robots.
Odyssey-3 from introduction to research preview, dated from Odyssey’s own posts and its two launch posts on X.
How Odyssey Says It Was Built
Odyssey says the training data combines internet video with event annotations, gameplay recordings with time-aligned keyboard and mouse inputs, and simulated rigid-body interactions. It then distils the model into a “few-step model capable of real-time interaction.”
Odyssey was founded in 2023, according to the 15 September post. Its news page lists a post titled “Our $310 Million Fundraise to Accelerate World Simulation”, dated 17 June 2026.
The Structural Read
The research preview and the benchmark claims answer different questions. The preview lets anyone prompt a world and move through it; the Physics-IQ and WorldMark numbers are Odyssey’s case that the worlds behave physically.
The machine demonstrations carry the commercial argument. In Odyssey’s account, the expensive part, broad knowledge of how the world moves, is learned once, and each robot arm, humanoid or car adds a small decoder or policy on its own data.
Odyssey states the limit itself: the WorldMark results measure properties of generated worlds, and a physical system still needs its own tests.
Odyssey, Meet Odyssey-3, 8 October 2026
“These results measure specific properties of generated worlds; applying the model to a physical system also requires evaluating the behaviors that matter for that machine and its tasks.”
Three Implications
ROBOTICS AND AUTONOMY TEAMS Odyssey invites physical-AI developers to get in touch; the posts we read give no price for developer access.
ANYONE COMPARING WORLD MODELS The WorldMark rankings are Odyssey’s own evaluation; Odyssey-3 leads three of the four splits it reports and is third in first-person real.
ANYONE TRYING THE PREVIEW The launch post on X says it is free; the posts we read do not say how long the free research preview runs.
The essay argues that what changes between AI paradigms is “where intelligence accumulates and who can capture it”, and names a Container Moat, “built by closed data loops inside a customer’s environment”, as “nascent but the deepest position in the stack.”
As we read it, Odyssey’s recipe splits along that line: a broad model trained on internet video, gameplay and simulation, then each machine’s own demonstrations to adapt it. The posts we read do not say who keeps those per-machine adaptations.
What Is Not Established
The posts we read do not give Odyssey-3’s parameter count, a price for developers, the size of the training data, or independent results beyond Odyssey’s own runs and evaluations. The WorldMark rankings are Odyssey’s evaluation, and the robot, humanoid and driving results are demonstrations Odyssey describes, not third-party tests.
The research preview is free, according to Odyssey’s launch post; the posts do not say how long the free preview runs or what it limits.
Business Engineer Framework
The Four Intelligence Moats
A Business Engineer essay on where intelligence accumulates in the AI stack, and who can capture it.
Odyssey has opened Odyssey-3 to the public as a free research preview and claims the top published score on Physics-IQ’s video-to-video test. Its case for the model rests on adapting one foundation to robot arms, humanoids and cars with tens of hours of data per machine, results that come, for now, from Odyssey’s own experiments.
A note on sourcing. We read Odyssey’s two posts on Odyssey-3, its 8 October launch post and its 15 September introduction, in full on 11 October 2026, along with its news page and the launch posts on X. The Business Wire copy of the announcement did not load for us. The benchmark scores, rankings and machine demonstrations are Odyssey’s own results and evaluations; we have not seen independent tests. Nothing here is a forecast, and nothing here is financial or investment advice.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur, HackerNoon and Search Engine People. Find him on LinkedIn and Substack. See how we source.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
