Odyssey-3 lets you walk through an AI world for free, and wants to drive robots next

Most AI models you use every day answer with text, images or short video clips. You ask, they deliver, and the conversation is over. A world model works differently. It does not just draw a picture of a place. It tries to predict how that place behaves: what happens when you take a step forward, turn your head, drop a ball or drive around a corner. And it keeps predicting, frame after frame, while you act inside it.

This week one of those systems became something anyone can try. The California AI lab Odyssey opened a free public research preview of Odyssey-3, its newest world model. You type a short description, the model builds an interactive environment in real time, and you can walk through it in first person or third person, move the camera on its own, and trigger events to see how the world reacts. Odyssey first showed the model on September 15 and published the full launch post, the benchmark results and the public preview on October 8, 2026.

The demo is the fun part. The bigger story is what Odyssey wants the same model to do next: steer robot arms, power humanoid robots, drive cars, fly drones and train other AI agents. Below is what is actually new, what the numbers really say, including an important caveat on the headline score, and why world models are turning into one of the most watched areas in AI.

The three things you need to know

• What is new. Odyssey opened a free research preview of Odyssey-3. You write a prompt, the model generates an interactive world in real time, and you can move through it and trigger events. The free demo runs on a faster variant called Odyssey-3 Flash. Developers who want API access have to apply through a contact form.
• The numbers. The base model has 14 billion parameters and generates video at 832 by 480 pixels. Odyssey-3 Pro runs at 1280 by 720. Pro reached 66.1 points on the Physics-IQ Verified video to video benchmark, the highest score on that leaderboard, but that result came from a single run that picked the best of eight generated videos per task. Averaged over four normal runs, Pro scored 63.37.
• The big plan. Odyssey says the same foundation model can control robot arms, humanoid robots, cars, drones and video game characters. A driving policy trained on about 20 hours of driving data, with the Odyssey-3 backbone kept frozen, drove a car on real roads in India, according to the company.

What a world model actually is

A large language model learns which word is likely to come next. A world model learns which moment is likely to come next. Odyssey describes Odyssey-3 as a learned dynamical system built as an autoregressive diffusion transformer. In plain words: it looks at the frames it has already produced plus your latest input, such as a step forward or a new event, and generates the next frames that fit. Then it repeats that loop again and again, fast enough that it feels live.

To learn how things move, the model was trained on three kinds of data, according to Odyssey. The first is internet video with time stamped descriptions of what happens in each clip. The second is gameplay recordings where every frame is paired with the keyboard and mouse input that caused it. The third is simulated physical interactions between solid objects, with captions and metadata. Together these connect what the world looks like with what happens next and, where possible, with the action that caused it.

Real time is the hard part. Diffusion models usually need many steps to produce a single good frame. Odyssey says it used a post training method that combines distribution matching and adversarial distillation to create a few step version of the model. That distilled version is what makes it possible to walk through a generated world without waiting.

What you can do in the free preview

The preview lives in the browser. You describe a scene, choose first person or third person, and start moving. You can also move the camera independently and introduce events while the world is being generated, for example something falling or appearing, and then watch how the model responds. That is the real test of a world model: not whether one frame looks pretty, but whether the world stays consistent and reacts in a believable way when you poke it.

It is important to be clear about what the preview is and is not. It is a research preview, not a finished product. There is no published price list for the API, no self service key and no open weights. Robotics, driving and drone control are shown in Odyssey's own demos, but they are not part of what the public can try today. The preview is about generating and exploring interactive environments.

The numbers, and the caveat

Odyssey leads its announcement with physics. Physics-IQ is a benchmark that asks models to continue videos of real physical experiments, covering fluid dynamics, optics, solid mechanics, magnetism and thermodynamics. The model's continuation is then compared with what actually happened in the real experiment. Odyssey says Odyssey-3 Pro set a new record on the Physics-IQ Verified video to video track with 66.1 points, and scored 54.7 on the image to video track.

The record deserves a closer look. As The Decoder pointed out, the 66.1 comes from a single test run in which a selection method picked one of eight generated videos for each task. The benchmark rules ask for four runs with a reported standard deviation for any record claim, and the 66.1 run does not meet that bar. Without the selection step, Odyssey-3 Pro averaged 63.37 points over four runs. Odyssey's own chart notes say the same thing in small print: scores average four runs, and the best of eight result uses one run. Both numbers are on the official leaderboard, and both were submitted by Odyssey itself.

On a second benchmark called WorldMark, which measures how well a model follows controls, how good the images look and whether the world stays consistent over time, Odyssey says its model ranks first in three of four categories in its own evaluation: 77.2 in first person stylized, 79.0 in third person real and 76.3 in third person stylized. In first person real it places third with 80.6 points. Again, these are the company's own runs, not independent tests.

None of this means the model is weak. It means the honest headline is "near the top of a hard physics benchmark, with a record that comes with conditions," not simply "best in the world."

One model for robots, cars, drones and games

The most ambitious part of the launch is the claim that one world model can serve as a shared foundation for many different machines. The idea is simple to explain. The world model learns general knowledge about how the physical world behaves. For each machine you then train a smaller, specialized piece, an action decoder or a policy, that translates the model's predictions into the controls that machine needs.

Odyssey shows several examples. With only tens of hours of robot demonstrations, a policy built on Odyssey-3 completed manipulation tasks such as pouring cereal into a bowl and closing a screw box, and it showed recovery behavior that was not in the demonstrations, like reorienting the gripper after a missed grasp. The Swiss humanoid robot company Flexion built control policies on Odyssey-3, and Odyssey says those policies held up better than the tested vision language action baselines when the lighting changed.

For driving, Odyssey kept the Odyssey-3 backbone frozen and trained a driving policy on about 20 hours of driving data. The policy uses the model's visual representations to predict waypoints ahead of the car and drove in closed loop on real roads in India. The company also adapted the model to generate matching views for three front cameras at once after just 100 training steps, which could be useful for creating training data for self driving systems.

The model also steered a drone through a virtual indoor space, dodging obstacles and flying to targets on command. And it played Grand Theft Auto V, steering vehicles and fighting enemies. According to The Decoder, a controller trained on roughly two hours of GTA footage moved a character on horseback in Red Dead Redemption 2 without extra training. Finally, Odyssey shows an agent that receives a goal in plain language and tries to complete it inside a world generated by Odyssey-3, which points at a future where AI agents learn by acting in simulated worlds instead of only reading text.

All of these are company demonstrations. They are interesting signals, but they have not yet been reproduced by outside researchers.

Who is behind Odyssey

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke, who previously spent about a decade building self driving cars. That background shows in the pitch: in autonomous driving, predicting what the world will do next is the core problem, and Odyssey wants to take that idea far beyond roads. In June 2026 the company raised 310 million US dollars in a Series B round, with investors including Amazon and AMD Ventures, according to the company and press reports.

Odyssey is not alone. Google DeepMind works on interactive worlds with Genie 3, Nvidia offers its Cosmos models for physical AI, and World Labs, the startup founded by Fei-Fei Li, builds similar technology. AMD announced in late September that it plans to buy World Labs for about 8.2 billion US dollars. World models have clearly moved from a research curiosity to a strategic bet for big chip and cloud companies.

Why it matters

Language models learned from the internet's text, and they became very good at talking about the world. But robots, cars and drones do not need a good essay. They need to know what happens if they push, turn or brake. Collecting real world data for every robot and every situation is slow and expensive. A strong world model could act as a shared simulator and a shared foundation, so each new machine needs far less of its own training data. Odyssey's examples, tens of hours for robot arms and about 20 hours for driving, are exactly the kind of numbers that would make that bet pay off if they hold up.

There is a second reason this matters. Free public access changes how a model gets judged. Until now, many frontier world models were visible only in polished company videos. With a public preview, anyone can try to break Odyssey-3: walk too far, trigger strange events and check whether the world stays consistent. That pressure from real users usually tells us more than any launch chart.

Dany's take

I like this launch, and I like that people can try it instead of only watching a highlight reel. Walking through a world that an AI builds while you move is the kind of demo that makes a hard research idea easy to feel. At the same time I would read the benchmark headline carefully. A record built on picking the best of eight tries is not the same as a normal run, and the real average of 63.4 is the number I would quote. The robot, driving and drone demos are exciting, but they are still the company's own results. If independent tests in the next weeks confirm even part of it, world models will be one of the biggest AI stories of the coming year. Have you tried the preview yet? I would love to hear what kind of world you built.

Sources

• Odyssey launch post: Meet Odyssey-3: Our Most Powerful Foundation World Model
• Odyssey background post: Introducing Odyssey-3
• The Decoder: Odyssey-3 is a new generative world model that you can try for free