Genie 4: what Google DeepMind's world model does, and what it means for business

On 27 October 2026 Google DeepMind announced Genie 4, a model that generates a world you can move through in real time, with sound, and change as you go. "Today, we are announcing Genie 4, our new world model," the DeepMind announcement begins, and it calls it "our first real-time world model to generate audio natively." For now it is a research preview: Google says it has "started rolling out Genie 4 to a limited group of trusted testers".

"Genie 4" has been climbing Google Trends since, and the demo videos are striking. This piece sets out what the model actually does, what a world model is and why labs spend so much on them, where the announcement deserves scepticism, and what, realistically, it means for a business. The short version of the last part: most businesses will not use Genie 4 this year, and that is fine.

What Google DeepMind announced with Genie 4

Genie 4 takes a text prompt and produces an interactive scene that responds as you act in it. The DeepMind post and Google's own blog describe four main things.

Real-time control. You move through the world with ordinary keyboard controls, and you can type instructions while it runs. Google's examples are changing the camera angle or the weather, or adding a character. TechCrunch lists adding people or objects among the free-form edits.

Native audio. In Google's words, "Every world also generates native audio on the fly, including ambient sounds, music, and interactive dialogue." The sound is generated with the picture rather than added afterwards.

Gemini as the planner. Genie draws the frames, but other models keep the scene on track. DeepMind says Gemini Omni anchors appearance and behaviour to the real world, while Gemini 3.8 Flash acts as a guiding agent that plans steps, checks progress and adapts when the user intervenes. TechCrunch describes Gemini as "acting as the director that plans steps and checks frames while Genie creates the visuals."

More realistic physics and people. DeepMind points to simulated aerodynamics and collisions, and to characters that move naturally and can be steered in real time. In one research demo, a Gemini 4 Argon model broke a task into steps and Genie 4 simulated Apptronik's Apollo humanoid robot playing checkers.

What a world model is, and why labs care

A language model predicts the next word. A world model predicts what happens next in a place. When DeepMind launched Genie 3 in August 2025, it defined world models as "AI systems that can use their understanding of the world to simulate aspects of it": given what the world looks like now and what you do, the model shows you what it would look like next. Nobody writes the physics rules or builds the 3D level by hand. The model has learnt, from a very large amount of video, roughly how things tend to behave.

Labs care about this for a few reasons.

Training robots and agents. Real-world robot training is slow, expensive and occasionally dangerous. A simulator that can produce endless varied scenes, a kitchen, a warehouse, a street in the rain, lets a robot or software agent practise far more than it could in reality. The Genie 3 post talks about training agents in "an unlimited curriculum of rich simulation environments", and Google frames Genie 4 as "enabling safe training environments for robots and agents".

Games and interactive media. A world you can walk through, generated from a sentence, is an obvious prototype tool for game designers.

Film and pre-visualisation. Directors already block out scenes roughly before they shoot. Google lists filmmaking, journalism, gaming and manufacturing among the industries it has in mind. DeepMind engineer Diego Rivas told TechCrunch that testers found it useful for film prototyping, "in the ideation phase, in the storyboarding phase — anytime you're kind of in need of rapid ideation."

The long game. DeepMind's Genie 3 post called world models "a key stepping stone on the path to AGI." That explains the spending; it does not make the current model useful to you.

Genie 4 vs Genie 3: what changed

Most of the changes are in what the world contains and how you steer it, not in how long it lasts.

Sources: Genie 3 announcement, Genie 4 announcement, DeepMind's Genie model page and TechCrunch.

The session length needs a note, because the reporting disagrees. DeepMind's model page says "The model can support a few minutes of continuous interaction, rather than extended sessions", nearly the same wording it used for Genie 3. TechCrunch reports that Genie 4 "can maintain a consistent scene for up to two minutes, double its predecessor's limit." For comparison, when Google opened Project Genie, its public tool built on Genie 3, to US AI Ultra subscribers in January 2026, each session was capped at 60 seconds. Whichever figure you take, the unit is minutes, not hours.

Who can use it, and what it costs

According to DeepMind, "Genie 4 is available as a research preview for trusted testers." Broader access is planned in two places: Google AI Ultra subscribers through Project Genie, and early-access partners of the Gemini API and Google AI Studio. TechCrunch puts that rollout "in the coming weeks"; Google's own post says "coming soon". There is a waitlist for Project Genie.

Google has not published a price for Genie 4. Figures for API pricing are circulating online, but none of them comes from Google, and DataCamp's review of the launch reaches the same conclusion: "Genie 4 has no published price." Until Google publishes one, treat any per-minute or per-session number you see as a guess.

Where to be sceptical about Genie 4

The model is impressive. The announcement is also a launch, and should be read as one.

DeepMind lists the limits itself. Its post says "visual details and object permanence can still shift or drift over time", especially in longer runs, after several edits, or while moving through a scene. It names complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as open research problems. For a robot simulator, a world that forgets where it put the cup is a serious flaw, not a cosmetic one.

Demos are not a product. The demo page notes "Sequences shortened" and calls results "for illustrative purposes." There is no benchmark table and, as DataCamp put it shortly after launch, "Nobody outside Google has evaluated it." When TechCrunch tried Genie 3 through Project Genie, its reviewer found that it "tended to fail when it came to photorealistic or cinematic worlds", that controls were often unresponsive, and that characters sometimes walked "right through walls or other solid objects." Genie 4 may well be better. Nobody independent has shown it yet.

Google says it is early. Rivas told TechCrunch "it's still early days for Genie 4", and that it is not yet clear what the end-to-end product will look like. When Project Genie launched, DeepMind's Shlomi Fruchter said the team did not see it as "an end-to-end product" people would use daily.

Compute is the unspoken constraint. Generating every frame and its sound live is expensive. Google has not disclosed what Genie 4 costs to run, but Fruchter explained Project Genie's 60-second limit with: "we wanted to bring it to more users." Short sessions are, in part, rationing. That matters if you were imagining running thousands of simulated hours to train something.

There is a commercial race on. TechCrunch notes Runway released a playable world model with audio a couple of months earlier, and that AMD is acquiring Fei-Fei Li's World Labs. Google has every reason to show its most polished demo, which is why independent testing matters.

What Genie 4 means for a business

For almost every business, the honest answer for 2026 is: nothing you need to act on. You cannot get production access, there is no price, and the model's own makers say it drifts after a few minutes. That does not make it irrelevant. It makes it something to understand now and use later.

Don't build on a preview. A research preview can change, slow down, get more expensive or be withdrawn. If a product or process depends on Genie 4 before it has a stable API, a published price and terms you can read, you are carrying risk you cannot price. Experiment if you get access; don't put it in anything customers rely on.

Know which uses are near and which are far. The nearest is creative ideation: pitch visuals, storyboards, rough walkthroughs of a space or a game idea, where a minute of drifting detail does no harm. Further off is training content, such as scenarios for safety or customer-service training, which needs the world to stay accurate and the text in it to be legible, both listed as unsolved. Furthest off, for most companies, is training robots or agents in generated worlds and using the output as synthetic training data. That is where the labs' real interest lies, and it depends on exactly the long-run consistency the model does not yet have.

Ask where simulation would help you, without Genie. The useful question is not "how do we use Genie 4?" but "where would we benefit from testing something before doing it for real?" That might be a warehouse layout, a new process, or an AI agent that should be tried against realistic cases before it touches live systems. Conventional simulation tools and plain test environments answer many of these questions today.

Test agents before they act. The idea behind world models, practise in a safe copy before acting for real, is good practice for any AI agent today. Agents in live systems carry risks that are easy to miss; our piece on AI agents and the CPU cache covert channel is one example.

Keep the bar where it belongs. A world model is a means, not a goal. If you are unsure whether AI belongs in a given process at all, our guide on whether your business really needs AI is a better place to start than a demo reel.

Genie 4 is a real step: sound, steering and more convincing physics in a world built from a sentence. It is also a research preview that holds together for minutes, with no price and no independent tests. Watch it, try it if you are offered access, and wait for the boring details, an API, a price and a few months of other people's results, before you build on it.