Google DeepMind unveiled Genie 4 on October 27, 2026, and the pitch is simple enough to state in one sentence: type a prompt, step into the world it makes, and then keep rewriting that world while you are still standing in it. The new Google Genie 4 world model generates photorealistic scenes in real time, and for the first time in the Genie line, those scenes come with sound of their own.
That last part is the headline change. Genie 3, released as a research preview in August 2025, could build a place you could walk through. It could not make that place sound like anything. Genie 4 can. Music, wind, waves, footsteps, and characters who talk back are generated alongside the visuals, not layered on afterward.
Don’t mistake this for a video generator with a new coat of paint. A video model renders a clip and you watch it. A world model simulates a place and you move through it. Press a key and the camera moves. Type a new instruction and the weather turns, or a stranger walks into the scene. The model has to predict what happens next, frame after frame, while you keep interfering.
Google Genie 4 world model: what actually changed
Four upgrades separate Genie 4 from its predecessor, and each one is easy to check against what DeepMind and independent coverage actually said.
First, sound. “Genie 4 is our first real-time world model to generate audio natively,” DeepMind said in its announcement, a line carried by Analytics India Magazine. That means ambient noise, music, and interactive dialogue are produced together with the world, so a character’s words can react to what you just did.
Second, steering. You navigate with keyboard controls and type plain-language prompts mid-simulation. Want rain? Type it. Want the camera somewhere else? Type that too. Gemini, Google’s AI model, adapts your request to fit the scene already running, so an edit lands as part of the world rather than as a hard cut to a new one.
Third, staying power. A scene now holds together for up to about two minutes, according to TechCrunch and The Verge, roughly double Genie 3’s limit of about one minute. Two minutes still sounds short. In world-model terms it matters, because drift is the enemy: the longer a simulation runs, the more details slip, objects change, and the scene quietly stops matching itself.
Fourth, people. DeepMind calls its simulated characters the biggest leap in this version. They move more naturally, respond to their surroundings, and can be steered in real time. Better physics helps too. DeepMind says Genie 4 handles things like aerodynamics and collisions more convincingly than before, which is the kind of detail that only shows up when something in the scene actually bumps into something else.
Gemini plays director
Genie 4 does not work alone. Gemini sits inside the simulation loop, and the division of labor is worth understanding because it explains where Google thinks this technology is headed.
Gemini Omni anchors how things look and behave, giving the world a consistent visual and physical character. Gemini 3.8 Flash acts as a guiding agent, watching each frame, planning the next steps, and checking that those steps actually happened. When you type a mid-scene edit, it is Gemini’s job to work out what that edit should mean inside the world you are already in.
One research experiment shows why that pairing matters. In work tied to Gemini Robotics, a Gemini 4 Argon model broke a task into steps while Genie 4 simulated them in real time, including a simulated humanoid robot playing checkers inside a generated world. TechCrunch reported the experiment as a possible way to test robot scenarios in simulation before they happen in the real world. Apptronik’s Apollo robot was the machine in the demo, rehearsing inside a place that only exists as generated video and audio.
Google research director Tejas Kulkarni put the direction of travel plainly: “Going forward, there will be more hybrid world models where Gemini writes code and text and Genie receives it in real time to simulate reality across different levels of abstraction.” That is not a product promise. It is a research bet that language models and world models belong in the same loop.
Why robots and game studios care first
DeepMind is pitching Genie 4 at film pre-visualization, game prototyping, interactive learning, and robot and agent training. Those four uses share one trait: they all benefit from a cheap place to fail.
A game developer can go from an idea to an explorable prototype before writing engine code. A film team can block out a scene and hear how it might sound before booking a set. A robotics team can run a scenario dozens of times in simulation, including the awkward edge cases, before a physical machine ever moves. None of that replaces the real thing. All of it makes the real thing cheaper to attempt.
Investors are watching the same logic. TechCrunch noted that the world-model field is starting to consolidate, with Fei-Fei Li’s World Labs being acquired by AMD, Amazon in talks to buy Decart, and Rivian acquiring Third Dimension to bring simulation in-house for self-driving tests. Runway released its second world model, GWM Worlds 2, with interactive video and audio a couple of months before Genie 4. Google is not entering an empty room. It is trying to stay ahead in one that is filling fast.
What you can’t do yet
Honesty requires the other half of the story, and DeepMind does not hide it. Genie 4 is a research preview, open first to trusted testers. Google says it is coming soon to Google AI Ultra subscribers through Project Genie and to early-access partners of the Gemini API and Google AI Studio, with a waitlist open for everyone else. No price for Genie 4 itself has been announced, and no date for full public access has been given.
Limits remain baked in. Details and objects can drift over longer sessions, especially after repeated prompt edits. Audio is not always accurate. Text inside the world is not always legible. A two-minute scene is a step forward, but it is not a game level, and it is not a film. Anyone claiming to sell you Genie 4 access today is selling something Google has not released.
DeepMind engineer Diego Rivas framed the current value the same way, telling reporters that real-time features deliver the greatest value when rapid ideation is needed, such as during concept development or the storyboard stage. In other words, Genie 4 is most useful right now exactly where ideas are cheapest: at the start, when you want to see something quickly and are willing to throw most of it away.
Every frame and sound Genie 4 generates carries SynthID, Google’s hidden watermark for AI-made content. That detail matters as these worlds get more convincing, because the gap between a generated scene and a recorded one is about to get harder to spot with the naked eye.
Why this matters
World models have spent years as a research curiosity with a demo problem: impressive for thirty seconds, shaky by sixty. Genie 4 attacks the two weaknesses that kept them there, sound and staying power, while handing the steering wheel to the user mid-scene. None of those fixes is complete. Together they move the technology from something you watch to something you use, at least for short bursts.
Watch what happens next in three places. First, whether the waitlist turns into real access for Ultra subscribers and API partners, and how quickly. Second, whether robotics teams publish results from training in Genie worlds rather than just demos of it. Third, whether rival world models match the audio and steering combination or push further on duration, because two minutes is still a ceiling everyone in the field is bumping against.
Genie 4 will not replace games, films, or robot labs. It might change how the first draft of each one gets made, and that is a bigger shift than another few seconds of coherence suggests.
FAQ
What is Google Genie 4?
Genie 4 is Google DeepMind’s world model, announced October 27, 2026. It generates photorealistic, interactive virtual worlds in real time from a text or image prompt, complete with native audio, and lets you move through them with keyboard controls.
How is Genie 4 different from Genie 3?
Genie 4 adds native audio, live text steering mid-simulation, and about two minutes of scene coherence, double Genie 3’s roughly one minute. It also has better physics and more natural-looking characters, according to TechCrunch.
Can I try Google Genie 4 today?
Not yet, unless you are a trusted tester. Google says Genie 4 is coming soon to Google AI Ultra subscribers through Project Genie and to early-access partners of the Gemini API and Google AI Studio, with a waitlist open.
What is Google Genie 4 used for?
DeepMind pitches it for game prototyping, film pre-visualization, interactive learning, and training environments for robots and AI agents, where a scenario can be rehearsed in simulation before anything happens in the real world.
Sources: Google DeepMind, TechCrunch, The Verge