What is Odyssey-3?

Odyssey-3 belongs to the world model category: models that learn how an environment looks and behaves well enough to keep simulating it as you act inside it. Odyssey pitches it as one model for several jobs at once: interactive entertainment, AI agents, robots, vehicles and drones.

Key facts

Specs are from Odyssey's announcement as reported by The Decoder. Odyssey has not published a frame rate or latency figure for Odyssey-3.

What you can do in the free preview

The public preview runs Odyssey-3 Flash in the browser. You write a text prompt to generate a world, choose a first-person or third-person perspective, move through it, trigger events and watch the scene respond as it is generated. Nothing is pre-rendered: if you turn left, the model invents what is to your left.

The robot, vehicle and drone controllers Odyssey showed in its demos are not part of the public preview.

How it works

Odyssey-3 was trained on three kinds of data: internet video paired with descriptions of what happens in it, video game footage paired with the keyboard and mouse inputs that produced it, and simulated physical interactions. The game footage is what teaches it how an action maps to the next frame.

To make it fast enough to respond live, Odyssey used an extra training step that cuts the number of compute steps per frame. That is the same broad recipe other real-time video labs describe: generate frames one after another, and distill the slow multi-step sampling down to a few steps.

Benchmarks and demos (company-reported)

• ✓Physics-IQ: Odyssey reports 66.1 points for Odyssey-3 Pro on the video-to-video benchmark, from a single run in which one of eight generations was picked per task. Without that selection it averaged 63.37 across four runs. The Decoder notes the 66.1 result does not meet the benchmark's own rule of four runs with reported deviation.
• ✓WorldMark: first place in first-person stylized (77.2), third-person real (79.0) and third-person stylized (76.3); third in first-person real (80.6).
• ✓Games: an agent built on Odyssey-3 played GTA V, and a controller trained on about two hours of GTA footage moved a character on horseback in Red Dead Redemption 2 with no extra training.
• ✓Robots and drones: robotic arms trained on a few dozen hours of demonstrations, a humanoid-robot partnership with Flexion, and a drone controller trained in simulation.

Odyssey-3 vs related models

• ✓vs Project Genie (Google DeepMind): both generate explorable worlds from a prompt. Project Genie runs 60-second sessions at 720p and 24 fps and needs a US Google AI Ultra subscription; Odyssey-3's preview is free. See Genie 3.
• ✓vs Decart Oasis 3: Oasis 3 is aimed squarely at driving and robotics simulation and sold by API per second. Odyssey-3 targets both entertainment and physical AI.
• ✓vs Runway Solaris: Solaris generates software interfaces rather than places. See interface world model.
• ✓vs live camera tools: Odyssey-3 invents a world from text. A live video-to-video tool such as LiveGen starts from your real camera feed and transforms it. Same real-time principle, different input. See real-time interactive video.

For every 2026 release side by side, see the real-time AI video and world model tracker.

Limitations

• The preview generates interactive environments only; the robotics and vehicle capabilities shown in demos are not available to the public.
• Benchmark and demo results come from Odyssey's own evaluations and submissions.
• Base resolution is 832 × 480, below the 720p of Project Genie and Odyssey-3 Pro.