In earlier editions world models sat in the research-frontier chapter, a promising idea with demos. In 2026 they became products you can pay for.
A world modelWorld modelA model that learns an internal simulation of how a scene evolves, aiming to keep physics, objects, and space consistent over time. A frontier for long, coherent video., instead of generating a fixed clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward., generates an environment you move through. It predicts the next frame in response to your inputs, so it behaves like a playable space rather than a finished video.
Google DeepMind's Genie 3 is the research flagship. It generates interactive worlds in real time at 720p and roughly 20 to 24 frames per second, and it holds changes in memory for about a minute, enough that a space stays consistent as you look away and back. Two honest caveats keep this in proportion. It is a limited-access research prototype, not a shipping APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface., and it generates frame by frame as autoregressiveAutoregressiveGenerating output one piece at a time, each conditioned on what came before, the way language models write text. Some image and video models work this way instead of by diffusion. video rather than exportable 3D geometry. Its persistence is measured in minutes, not hours.
The commercial edge belongs to World Labs, Fei-Fei Li's spatial-intelligence company. Its product Marble, which launched commercially in late 2025 and reached general availability with paid tiers in early 2026, generates persistent, explorable 3D environments from a promptPromptThe text description you provide to a model to specify what you want it to generate., an image, or a video, and, crucially, exports them as Gaussian splats, meshes, or video that drop into real game and film pipelines. Paid plans run up to around ninety-five dollars a month for commercial rights.
Around them sits a well-funded field: NVIDIA's Cosmos as a world-model foundation for robotics and simulation, Decart's real-time playable models, and Odyssey among the startups. The through-line is that generation is moving from making a clip to making a place.
Check your understanding
pass: 5 of 7
Answer at least 5 of 7 correctly to unlock the next chapter.
1. What changed about world models in 2026?
2. How does a world model differ from a normal video model?
3. What are the specifications of Google DeepMind's Genie 3?
4. What are the two honest caveats about Genie 3?
5. What crucial capability does World Labs' product Marble offer that keeps it commercially ahead?
6. Who founded World Labs, the company behind Marble?
7. What moment should you watch for, when the two halves of the world-models category meet?
The weekly briefing
Get the week's moves in your inbox.
A short, sourced digest of what actually moved across generative AI, every week. Free.
Free. One email a week, no spam, unsubscribe anytime. Prefer a reader? RSS.