Google DeepMind introduced Genie 4 on October 27, 2026, a world model that builds interactive scenes in real time and generates the sound as well, according to DeepMind’s announcement. You type a prompt, move through the result with keyboard controls, and change the weather or add a character while it runs. That makes it less a video tool and more a sandbox. Almost nobody can try it yet, so the useful question is what Genie 4 is actually for and how far it is from your screen.
TL;DR: Genie 4 is DeepMind’s newest world model. It turns a text prompt into a navigable scene, lets you steer it with natural language, and generates ambient sound, music and dialogue alongside the visuals. It is a research preview for trusted testers now, with Google AI Ultra subscribers next through Project Genie. DeepMind states no price and lists several known limits.
What Genie 4 does that a video generator does not
A video generator hands you a finished clip. Genie 4 keeps the scene running and lets you push on it, which is a different kind of tool.
DeepMind says you can change the weather, move the camera, or add a person or object using plain language, and the world adapts as you go. The model also simulates things like aerodynamics and collisions, and characters move and react to their surroundings more naturally.
The audio is the headline addition. DeepMind describes Genie 4 as the first real-time world model to generate audio natively, covering ambient sound and dialogue that changes on the fly. Ask for a storm and the wind is supposed to arrive with it, which is the kind of detail silent demos never had to get right.
All of that depends on something keeping the scene coherent between your inputs.
How Gemini keeps a Genie 4 world on track
Genie 4 is built on the Gemini family. DeepMind says Gemini Omni anchors a simulation in real-world appearance and behavior, while other Gemini models handle the reasoning that keeps it continuous.
In the real-time loop, Gemini 3.8 Flash acts as a guide. It watches the scene, breaks a goal into steps, checks whether each step happened, and adjusts when the world shifts or you step in.
DeepMind also shows a robotics experiment that pairs Gemini 4 Argon and Gemini Robotics with Genie 4. Argon splits a task into steps and sends them to Genie 4 as text captions, and the demo has Apptronik’s Apollo humanoid playing checkers. It is a lab demo, not a product, but it shows why DeepMind frames world simulation as training ground for robots and agents.
The practical question is who gets to touch any of this first.
Who can use Genie 4 and when
For now, Genie 4 is a research preview for trusted testers. DeepMind says it is coming soon to Google AI Ultra subscribers through Project Genie, and to early access partners on the Gemini API and Google AI Studio. A waitlist for Project Genie is linked from the post.
| Access path | Status |
|---|---|
| Trusted testers | Research preview, available now |
| Google AI Ultra via Project Genie | Coming soon, waitlist open |
| Gemini API and Google AI Studio | Coming soon, early access partners |
| Pricing | Not stated |
DeepMind gives no dates and no price, so anyone planning around it is planning around a “soon.” Trusted testers named film pre-visualization, game prototyping, and turning single archival photos into simulated historical settings as early uses.
Those early uses make the limits worth reading closely.
Where Genie 4 still falls short
DeepMind is direct about the rough edges. Visual details and object permanence can drift, especially in longer sessions, after several prompt edits, or while navigating.
It also lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as active research areas. That is a modest list of promises for a model that sounds like science fiction, and it makes the announcement easier to believe.
On safety, DeepMind says it ran internal evaluations before the preview and that all generated video and audio carries SynthID watermarking. Apple recently added SynthID support in iOS 27.2, so the watermark is starting to show up in tools people already use.
Which brings the story back to what a normal reader should take from it.
What to watch before the Ultra rollout
DeepMind says a language model could plan a scene’s structure and story while a model like Genie renders how it looks and sounds, which it calls hybrid world models. That is the long game here, and Genie 4 is an early step toward it.
For now, the thing to watch is the Project Genie waitlist and whether longer sessions hold together once more people can test them. If the drift DeepMind admits to stays small outside a controlled preview, native audio could be the part that makes these worlds feel finished.
