Google Genie 4: Worlds You Can Talk To — and Steer Live

Type a sentence and a world appears. Walk into it. Then type another sentence — "make it rain" — and the weather changes while you are standing in it. That is the pitch behind Genie 4, the world model Google DeepMind announced on October 27, 2026. It is the fourth generation of a line that started with a research paper in early 2024, and it is the first version with a voice: the worlds it generates come with their own sound, music and talking characters, produced at the same time as the picture.

There is a catch, and it is a big one. Almost nobody can try it yet. Genie 4 is a research preview for a small group of trusted testers, with a waitlist open for everyone else. Google says Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio are next in line, with no date attached. There is no published price for Genie 4 itself, and no API rate at all. This article covers what the model actually does, what DeepMind itself says it cannot do yet, and what the only route in will cost you.

What Genie 4 actually is

A world model is not a video generator. A video model produces a clip and you watch it. A world model produces a place and you move through it: press a key and the camera moves, type a prompt and the world changes. Genie 4 generates explorable, photorealistic worlds in real time from a text prompt or an image, using standard keyboard controls for navigation and camera work.

DeepMind's announcement, written by researchers Raia Hadsell, Diego Rivas and Hang Qi, sets out four advances over Genie 3. First, sound. Genie 4 is, in DeepMind's words, its first real-time world model to generate audio natively: every world carries the ambient sound of its own setting, from wind in the trees to waves on a beach, plus music and interactive dialogue. Characters converse and react to what you do or to events in the world, in real time. DeepMind calls human speech "the true frontier" of this work.

Second, people and physics. The company calls simulating people "the most dramatic leap" in this release: characters that move naturally, respond to their surroundings, and can be steered as the scene runs. The model also simulates real-world physics, including aerodynamics and collisions, because it was trained on data that closely reflects the physical aspects of the world, according to DeepMind.

Third, live steering. You can change the world mid-simulation with natural-language prompts — change the weather, move the camera, add a person or an object — and Gemini adapts your prompt to fit the surrounding scene. Between your inputs, Gemini monitors and steers the scene to keep the world progressing coherently.

Fourth, events that play out. Set something in motion and Genie 4 simulates what happens next: how the world responds, how it sounds, and how the people in it move, react and speak. That is the difference between a generated clip and a simulation with cause and effect.

The second model in the loop

The part of the announcement that matters most for understanding what Genie 4 is — and is not — is the Gemini machinery wrapped around it. Genie 4 incorporates capabilities from the Gemini family: Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models provide the reasoning that keeps a simulation coherent.

In practice, Gemini 3.8 Flash serves as a guiding agent inside the loop. As Genie 4 generates a world in real time, Gemini 3.8 Flash continuously observes the scene, breaks a goal down into a sequence of steps, and reviews incoming frames to track progress and verify that each step succeeded. When the world shifts or you intervene, Gemini adapts its plan in the moment. DeepMind's own framing is that Gemini maps the structure and plans the narrative timeline, while Genie generates the immediate physical reality, pixel by pixel.

DeepMind also showed an experiment pairing Genie 4 with Gemini 4 Argon, its flagship model, in which Argon breaks a task down into steps — in the demo, Apptronik's Apollo humanoid playing a game of checkers — that Genie 4 renders in real time. The point of the experiment is robot training: test how an AI system navigates different environments in simulation before bringing it into the real world. That is also the most concrete near-term use DeepMind names for the technology, alongside new interactive experiences.

One number worth handling carefully: several independent reports, following TechCrunch, say Genie 4 can keep a scene consistent for up to about two minutes, roughly double Genie 3's limit of around one minute. DeepMind's own post publishes no duration figure at all, so treat the two-minute number as reported, not official. The same applies to resolution, frame rate, parameter count and benchmarks: Google published none of them.

What it costs — and the fine print

The honest summary is one sentence: there is no Genie 4 price, because there is no Genie 4 product yet. It is a research preview, and the only access route Google has named runs through a subscription you may already be paying for — or may not want to.

Google AI Ultra is the subscription that matters here, and its pricing has moved recently enough that the numbers online disagree. Google's original Ultra tier was $249.99 a month. Recent reporting describes a restructured lineup of two tiers — $99.99 a month for a 5x tier and $199.99 a month for a 20x tier — and at least one report ties Project Genie specifically to the higher tier. Google's own plans page did not render readable prices when we checked it on October 28, 2026, so treat the $99.99 and $199.99 figures as reported, and check the live plans page before subscribing on the strength of them.

Two more pricing facts are worth stating plainly. First, the guiding model inside Genie 4, Gemini 3.8 Flash, is a paid API model: pricing trackers list it at $0.75 per million input tokens and $3.75 per million output tokens on a promotional rate reported to rise on January 1, 2027. That is the cost of the reasoning loop, not of Genie 4, and it is reported rather than read from Google's own pricing page, which did not render a readable rate in our check. Second, if and when Genie 4 reaches the Gemini API, expect its own rate card — DeepMind has published nothing about one.

There is also a free tier of sorts: the waitlist for Genie 4 in Project Genie costs nothing to join. It buys you a place in line, not a date.

What trusted testers are actually doing with it

DeepMind has been working with cohorts of trusted testers across creative industries, journalism, education, gaming and engineering, and the announcement lists what they reported back. Filmmakers are using Genie 4 as a rapid draft mode for pre-visualization: step into a live simulation of a scene and run multiple takes in real time, adjusting camera angles, lighting, dialogue and action on the fly. Photojournalists are exploring how a single 2D archival photograph can become a grounded simulation of that environment in seconds, including visualizing how short-term events like a flash flood or wildfire unfold. Game developers are going from an idea to an explorable prototype in seconds, testing visual styles and interactive mechanics through natural-language prompts before writing engine code. Enterprise partners are evaluating it for rapid "zero-to-one" scenario generation.

None of that is a product yet. It is a research preview, and DeepMind's own language is careful: the model "remains an early, experimental model" and the company is "sharing Genie 4 as a preview to learn alongside our community." The tester stories are company-selected, and no independent hands-on review exists yet, because almost nobody outside Google's tester cohorts has used it.

What Genie 4 cannot do yet

To DeepMind's credit, the announcement names its limits in plain language, under a section titled "Understanding current limitations." Visual details and object permanence can shift or drift over time, especially as a simulation runs longer, after multiple complex prompt edits, or while navigating through the scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy and long-term consistency are all listed as active areas of ongoing research.

That drift is the central failure mode of world models, and it is why the two-minute consistency figure matters: the longer a simulation runs, the more details slip. Two minutes is long enough for a scene to develop — a character can cross a room, weather can turn, a conversation can go somewhere — but it is not a game level, a film scene or a training shift. It is a sketch that holds together for about as long as a short song.

There is also a safety layer worth noting. Genie 4 includes strict safety guardrails consistent with Google's policies for other models, and DeepMind says it ran internal safety evaluations before the preview. All video and audio generated by Genie 4 carries SynthID, Google's imperceptible watermarking technology, which marks the output as AI-generated across visual and audio modalities.

Genie 4 vs the field

Genie 4 does not arrive in an empty field. Runway's GWM Worlds 2 also responds to user inputs and includes audio, according to TechCrunch's reporting, so the feature gap between the two is narrower than the launch headlines suggest. NVIDIA's Cosmos family of world foundation models targets robot training rather than interactive consumer worlds, and NVIDIA sells no hosted Cosmos API — the models are downloadable. World Labs' Marble is another interactive-world effort. No verified head-to-head benchmark exists between any of them, so the honest comparison is about access and intent, not scores.

For the wider context on the robotics side of world models, see our NVIDIA Cosmos coverage. The through-line is the same in both stories: world models are being built first as simulators for agents and robots, and only second as something a person plays with.

Pros and cons

Pros: the first DeepMind world model with native audio and talking characters, generated in the same pass as the visuals; real-time text steering that actually changes a running world; a Gemini model in the loop that plans multi-step events and checks its own work; photorealism and human motion that testers describe as a genuine step up from Genie 3; a clear, named use case in robot and agent training; SynthID watermarking on all output.

Cons: almost nobody can use it — trusted testers only, with no date for wider access; no published price and no API rate, so the cost of the only access route is a reported subscription tier; scenes drift after about two minutes and details slip with longer runs or repeated edits; no published resolution, frame rate, parameter count or benchmark; the most impressive demos are company-selected, with no independent testing yet.

FAQ

What is Genie 4?

Genie 4 is Google DeepMind's real-time world model, announced October 27, 2026. It generates explorable, photorealistic worlds from a text prompt or an image, with native audio — music, ambient sound and interactive dialogue — generated at the same time as the visuals. You navigate with keyboard controls and steer the scene with text prompts while it runs.

Can I use Genie 4 today?

Almost certainly not. It is a research preview for trusted testers. Google says it is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. A waitlist is open for Project Genie; joining it costs nothing and guarantees no date.

How much does Genie 4 cost?

Google has published no price for Genie 4 and no API rate. The only confirmed route in is Google AI Ultra, a paid subscription: recent reporting describes restructured tiers of $99.99 and $199.99 a month, with Project Genie tied to the higher tier. Verify the current price on Google's plans page before subscribing — the figures are reported, not confirmed on the live page.

How is Genie 4 different from Genie 3?

Genie 3, previewed in August 2025, generated silent worlds capped at about a minute of consistency. Genie 4 adds native audio and spoken dialogue, more realistic people and physics, live natural-language steering mid-simulation, and about two minutes of scene coherence according to TechCrunch's reporting — roughly double. It also puts a Gemini model, 3.8 Flash, in the loop as a guiding agent that plans and checks multi-step events.

How long can a Genie 4 scene last?

TechCrunch reported that the model can hold a consistent scene for up to about two minutes, double Genie 3's limit. DeepMind's own announcement publishes no duration figure, so treat two minutes as reported. DeepMind does say that details and object permanence can drift over longer sessions, after multiple complex prompt edits, or while navigating.

Key takeaways

• Genie 4 is Google DeepMind's new world model, announced October 27, 2026, and it generates interactive worlds in real time with native audio, music and dialogue.
• It is a research preview for trusted testers only; Google AI Ultra subscribers in Project Genie and early-access Gemini API partners are next, with no date given.
• There is no published price for Genie 4, and the only route in runs through a paid Google AI Ultra subscription, reported at $99.99 to $199.99 a month.
• A Gemini model, 3.8 Flash, works inside the loop as a guiding agent, planning multi-step events and checking each step as the world renders.
• Scenes hold together for about two minutes according to TechCrunch's reporting, then details drift — DeepMind names that drift, along with legible text and audio accuracy, as open problems.

Checked and written on October 28, 2026, the day after DeepMind's announcement, when Genie 4 was a research preview with no public price, no API rate and no published benchmarks. The launch film looks like magic. The honest version is narrower and more interesting: a world you can talk to, for about two minutes, that almost nobody can try yet.

Sources: Google DeepMind, "Genie 4: A new generation of world simulation" (October 27, 2026); Google AI Ultra subscription reporting; TechCrunch reporting on scene consistency; Gemini API pricing documentation. All vendor claims in this article are attributed to DeepMind or to the reporting outlet named.