Skip to content

DirectoryTools

Odyssey-3 logo

Odyssey-3

Odyssey's foundation world model for robots, cars and games, not yet released.

Visit Odyssey-3

VideoPaidWe checked the prices on odyssey.systems on 29 Sep 2026.

Odyssey-3 homepage screenshot

Recaply review

Odyssey-3 review: one world model for robots and games, not out yet

Odyssey-3 is Odyssey's foundation world model, built so one pretrained model can underpin robots, humanoids, cars, drones and game-playing agents; Odyssey has shown research results but not released it.

20 hours
Of simulated driving data taught a policy built on it to drive a real car in India
77%
As far between safety-driver interventions as a policy trained on real footage
2 hours
Of GTA V footage, roughly, gave a policy that also rode a horse in Red Dead Redemption 2
1
Pretrained model behind Odyssey's robot, humanoid, car, drone and game tests

What is Odyssey-3?

Odyssey-3 is a foundation world model from Odyssey: an autoregressive diffusion transformer trained on a large body of video to learn physics, motion and cause and effect. Rather than being used only to generate video, it acts as a shared base for control. A small action decoder or policy, trained on a few to tens of hours of task data, turns what it has learned into commands for robots, cars, drones or game characters.

It is not available yet. Odyssey has published research results and says it will release Odyssey-3 publicly, but gives no date, price or access details and asks interested teams to get in touch.

Is Odyssey-3 available yet?

No. Odyssey has shared research results and says it will release Odyssey-3 publicly, but there is no API, demo, price or date yet.

Public release
Promised without a date; the announcement ends with a contact link rather than a sign-up.
Research partners
Poke & Wiggle is testing it across robot bodies and controls, and Flexion builds humanoid control on it.
Odyssey's API today
Serves Odyssey-2 Pro, with Odyssey-2 Max rolling out to existing users.
Price
Not announced.

Who is Odyssey-3 best for?

Odyssey-3 is best for robotics, driving, drone and game-agent research teams who want to follow Odyssey's results and ask for access once it opens.

  • Robotics research teams

    Odyssey reports robot arms learning tasks from tens of hours of demonstrations, including recovering from a missed grasp.

  • Driving and drone research

    Policies built on it drove a real car after 20 hours of simulated data and flew a simulated drone around obstacles.

  • Game-playing agent research

    Policies trained on GTA V play moved characters in two other games without extra training.

How much does Odyssey-3 cost?

Odyssey-3 has no price yet: Odyssey has not released it or said how it will be sold.

Not releasedNo price yetAccess and pricing not yet announced
  • Odyssey says a public release will follow
  • Interested teams can contact Odyssey

Odyssey has not said how Odyssey-3 will be sold. Its current models are offered through an API with fees set in a contract, and Odyssey-2 Max is in private beta. Full pricing on odyssey.systems. We checked these prices on .

How do readers rate Odyssey-3?

No one has rated Odyssey-3 yet. Be the first.

No ratings yetThe score and its stars show once 3 readers have rated Odyssey-3.
Your rating

Sign in to rate. One rating per account.

Your rating of Odyssey-3

Odyssey-3 pros and cons

Odyssey's results suggest one model can learn many kinds of control from little data, but nothing is released and nobody outside Odyssey and its partners has tested it.

Pros

  • One pretrained model underpinned policies for robot arms, humanoids, a car, a drone and game characters in Odyssey's tests.
  • Needs little task data by Odyssey's account: 20 hours of simulated driving was enough to drive a real car, and tens of hours of demonstrations taught robot arms new tasks.
  • Odyssey reports behaviour its training data never showed, such as a robot reorienting its gripper after a missed grasp and a GTA-trained policy riding a horse in another game.

Cons

  • Not released: there is no API, demo, download, price or date.
  • Every result so far comes from Odyssey and its partners; testing across robots with Poke & Wiggle is still under way.
  • It is not a ready-made controller: each robot, vehicle or game needs its own action decoder or policy trained on task data.
  • Its sim-trained driving policy went only about 77% as far between safety-driver interventions as one trained on real footage, and Odyssey gives no absolute distance.

What are the best Odyssey-3 alternatives?

Lyra 2.0 for downloadable robot-simulation worlds, World Labs Marble for 3D worlds you can use now, and Odyssey-2 for a free look at Odyssey's live video.

Questions about Odyssey-3

What is Odyssey-3?

Odyssey-3 is Odyssey's foundation world model, built so one pretrained model can underpin robots, humanoids, cars, drones and game-playing agents; Odyssey has shown research results but not released it.

Who is Odyssey-3 for?

Odyssey-3 is best for robotics, driving, drone and game-agent research teams who want to follow Odyssey's results and ask for access once it opens.

Is Odyssey-3 free?

Odyssey has not said. Odyssey-3 is not released, and Odyssey has published no price or access plan for it. Its current API models are sold under contracts, with fees billed monthly in advance.

What are the best Odyssey-3 alternatives?

The closest alternatives on Recaply are Lyra 2.0, World Labs Marble and Odyssey-2.

Where can I try Odyssey-3?

Nowhere yet. Odyssey's announcement shows research videos but offers no demo, API or download; it points interested teams to its contact form and says a public release will follow.

How is Odyssey-3 different from Odyssey-2 Max?

Odyssey-2 Max is a real-time world simulator you steer with prompts or actions, in private beta now. Odyssey-3 is a foundation model meant to sit under other systems: small policies trained on top of it use what it learned to control robots, cars, drones and game characters.

Can Odyssey-3 control a real robot?

In Odyssey's research, yes. With tens of hours of demonstrations, robot arms learned tasks such as pouring cereal and closing a screw box, and Flexion built humanoid control on it. Each machine needs its own trained action decoder or policy.

Our verdict

Odyssey-3 is one to watch, not one to use: Odyssey has shown promising research but released nothing yet.

In Odyssey's tests one pretrained model underpinned robot, humanoid, car, drone and game policies with little task data, yet there is no API, demo, price or date, and every result is Odyssey's or a partner's.

Skip it for
Anyone who needs a model to build on today.
Check first
When it launches, check who can get access, what it costs, and whether independent tests match Odyssey's results.