The GPT-2 Moment for World Models Is Here
Introducing the Odyssey API and Odyssey-2 Pro, enabling developers to build incredible new applications with frontier world models
Oliver Cameron
Today is a big day at Odyssey: we’ve released Odyssey-2 Pro—our most powerful world model yet—and launched a brand new developer API. We believe this is the GPT-2 moment for general-purpose world models, where weird and wonderful consumer, enterprise, and intelligence applications can now be explored. Developers all around the world can now head to our developer portal to begin integrating open-ended, interactive simulations into your applications.
When GPT-2 and the OpenAI API were released, developers built tools, side projects, early products, and ran research experiments. Through that exploration, the field began to understand what language models were incredible at, where they needed improvement, and what new applications might emerge. This moment is now here for world models.
Introducing Odyssey-2 Pro
We launched Odyssey-2 in October 2025, and it demonstrated that a world model can learn to simulate basic physics, dynamics, behaviors, and more, entirely from a large corpus of general video and interaction data. Odyssey-2 Pro improves upon it across every dimension, and is our most intelligent, interactive, and general-purpose world model yet, and the first available in our API.
A burgeoning world simulator
A world simulator—like Odyssey-2 Pro—is a model trained to predict how the world evolves over time, frame-by-frame, using large amounts of video and interaction data. Our model outputs continuous simulations that can be interacted with, with all examples you see below generated and streamed in real-time. Today Odyssey-2 Pro can run for minutes, and over time we intend to enable hours, days, and eventually years of simulation.
Every simulation was generated in real-time
Improved capabilities across the board
Odyssey-2 Pro is a significantly larger model, with more capacity to internalize how the world works from an even larger corpus of video and interaction data. This increased capacity—while still real-time—has resulted in improved physics, better dynamics, more authentic behaviors, and more realistic pixels. In addition, the model now streams 720P video at 22 FPS, meaning simulations feel sharper and more realistic. Lastly, interactions are faster and richer.
Odyssey-2 Pro is a big leap from Odyssey-2, and we’re excited to see these results emerge from our research and scaling efforts. Simple, causal prediction objectives have already produced surprisingly general intelligence in language models, and we see no reason this won’t hold true for world models. Going forward, we believe scale—similar to language models—is a major lever for performance.
Learning cause and effect
Authentic human behaviors
Improved visual quality
Responsive, rich interactivity
Introducing the Odyssey API
Odyssey-2 Pro outputs continuous, interactive simulations that can now be embedded into your application in multiple ways. We’ve built three API endpoints—interactive streams, viewable streams, and simulations—to support both online, real-time generation and offline, batch workflows. Together, they enable builders to integrate world models in many ways.
Generate simulations with user-specified actions at precise time steps. You provide a prompt, a set of actions, quality settings, a target simulation length, and the model returns a video
Embed an interactive stream of generated video into your application. At any moment, you can programmatically interact with it in real-time
Distribute a single interactive stream to many users
Today we're releasing our JavaScript and Python SDKs, with iOS and Android SDKs coming soon.
Integrate in 5 minutes or less
It's really simple to integrate a frontier world model into an existing or new application. Watch just how easy it is in this video.
It's time to build with world models
Odyssey-2 Pro and the Odyssey API are now available to any developer. Head to our developer portal, generate an API key, and integrate an interactive stream or simulation into your application in under ten lines of code. We see all kinds of new gaming, simulation, education, healthcare, companionship, support, training, and intelligence applications waiting to be built.
With frontier world models like Odyssey-2 Pro reaching this level of performance, we believe this is the same inflection point we saw with GPT-2 in language. Although these models are still nascent, large-scale experimentation can now begin, and developers can discover what’s possible.