Google Genie 4 is DeepMind's real-time world model, announced October 27, 2026 as a research preview for trusted testers. It generates a playable world as you move through it, with native music, ambient sound and character dialogue, and you can change the scene with text prompts mid-session. It is not publicly available yet: Google AI Ultra subscribers in Project Genie and early-access Gemini API partners come next.
Google Genie 4 is Google DeepMind’s newest world model: a neural network that generates an interactive, explorable world in real time from a text description. It was announced on October 27, 2026, and today it exists only as a research preview for trusted testers.
This explainer covers what Genie 4 does, how it differs from Genie 3, where it still breaks and who can get in. Every fact comes from DeepMind’s pages or press coverage linked below. It’s for creators, developers and anyone trying to work out whether world models matter to them yet.
TL;DR: what Genie 4 is
Genie 4 is a world model, not a video clip generator. You describe a place, step into it with keyboard and mouse, and the model draws each frame as you move.
It’s new in four ways:
- Native audio. DeepMind says it is its first real-time world model to generate audio natively: music, ambient sound and characters who talk back.
- Prompt steering. You type to change the weather, move the camera or add a person or object while the world keeps running.
- Longer scenes. TechCrunch reports up to two minutes of consistent scene, double Genie 3.
- Gemini in the loop. Gemini models anchor the look and keep the world coherent.
You can’t use it yet unless you’re a trusted tester.
What is Google Genie 4?
Genie 4 is the fourth step in DeepMind’s Genie line of world models, the successor to Genie 3. A world model predicts what a scene should look like next, given what you just did, so the environment is generated rather than authored.
DeepMind’s Genie page describes it as a research preview for trusted testers, with sessions of a few minutes. You navigate with standard keyboard and mouse controls, and text prompts steer the scene as it evolves.
The Genie 4 page on DeepMind’s site, where the trusted-tester waitlist link lives.
That makes it different from a game, which someone built ahead of time. It’s also different from a video model, which hands you a finished clip. For the wider field, see our guide to the best AI world models.
How Genie 4 works
Genie 4 generates each frame from the world so far plus your latest input. There is no stored level and no 3D asset library behind it.
DeepMind says Genie 4 builds on the Gemini family. Gemini Omni anchors the simulation in real-world appearance and behavior, and other Gemini models supply the reasoning that keeps the world coherent. Between your inputs, Gemini monitors the scene and steers it so it keeps progressing.
TechCrunch describes Gemini as a director that plans steps and checks frames while Genie draws the visuals. DeepMind has not published a parameter count, resolution or frame rate for Genie 4, so ignore any article that quotes one.
Native music, ambience and dialogue
Genie 4 generates sound in the same pass as the picture. Per DeepMind, that covers music, ambient noise and characters who hold real-time conversations that adapt to what you do.
Genie 3 produced silent worlds. Sound is the biggest change in this release.
DeepMind is open about the limit: real-time audio accuracy is still an area of ongoing research. Expect plausible sound, not guaranteed sound.
Steering a world with prompts mid-simulation
You change a running world by typing. DeepMind’s examples include altering the weather, moving the camera and adding a person or object.
Gemini adapts each prompt to fit the scene that already exists, so a new character appears in the street you’re standing on instead of resetting the world. This is the feature that separates Genie from a clip generator: you direct it live.
Edits cost consistency. DeepMind says visual details and object permanence can drift, especially after multiple prompt edits.
How long a Genie 4 world stays coherent
TechCrunch reports a consistent scene can last up to two minutes, double Genie 3’s limit. DeepMind’s own pages are less specific: the model “can maintain coherent environments over short sessions” and supports “a few minutes of continuous interaction.”
Treat two minutes as the press-reported ceiling, not a published spec. For comparison, DeepMind’s Genie 3 announcement listed 720p, 24 frames per second and consistency for a few minutes.
What Genie 4 still gets wrong
DeepMind lists the open problems itself:
- Drift. Visual details and object permanence can shift over time, especially in longer runs or while moving through a scene.
- Real places. Genie 4 can’t simulate real-world locations with perfect accuracy.
- Text. Legible text mostly appears only when it was in your world description.
- Physics. Complex physical interactions are still unsolved.
- Multiple agents. Modeling several users or independent agents in a shared world remains open research.
Google also told TechCrunch it is early and the end-to-end product experience is unclear. All generated video and audio carries SynthID watermarking, per DeepMind.
What Genie 4 is for
DeepMind’s own pitch is gaming, interactive learning and robotics. In one robotics demo, a Gemini model splits a task into steps and Genie 4 renders them in real time, with Apptronik’s Apollo humanoid playing checkers. The idea is to test how AI systems handle multi-step tasks before they run on real hardware.
TechCrunch adds that trusted testers see promise in film prototyping, such as ideation and storyboarding. A Google contact told it that turning those prototypes into finished creative assets is still a way off.
Who can access Genie 4 and how
Genie 4 is available now only to trusted testers. DeepMind says broader access comes next to two groups:
- Google AI Ultra subscribers in Project Genie, Google’s tool for generating interactive worlds from text or image prompts. TechCrunch says this is “in the coming weeks.”
- Early-access partners of the Gemini API and Google AI Studio.
There is a waitlist link on DeepMind’s Genie page. No price, rollout date or general API availability has been published, so don’t plan a project around a date yet.
Genie 4 vs video models like Veo and Sora
A video model renders a clip you then watch. Genie 4 renders the next frame in response to you, so it runs live and you steer it.
That trade shows in the limits. Video models give you longer, higher-resolution, finished footage you can edit. Genie 4 gives you interaction for about two minutes. If you need a clip you can drop in a timeline today, you want a video model, and our Google Veo alternatives and best open-source AI video models guides cover them. OpenAI’s Sora is closed, as our Sora alternatives guide explains.
For a close look at one of today’s clip models, see our Hailuo 3 review.
Should you care about Genie 4 now?
If you make video, not yet. You can’t get in, it’s built for interaction rather than finished clips, and sessions reportedly hold together for about two minutes.
If you build games, simulations or robotics tools, watch it closely. Audio plus live steering is a real step beyond Genie 3, and the Gemini API path means developers may get hands-on access first.
Conclusion
Genie 4 is the clearest sign yet that generated worlds are moving from demo to product. It adds sound, live steering and longer scenes, and it still drifts, fumbles text and ships to a handful of testers.
Watch Project Genie and the Gemini API for the first public access. Until then, clip models are the practical way to make AI video.