Google Genie 4: AI Worlds You Can Walk Through

Updated: October 2026 · 6 min read · By the CreatorSprout team

Launch coverage based on Google DeepMind's announcement and independent launch reporting — we have not tested Genie 4 hands-on, and it is not publicly available yet. Vendor claims and reported details are kept separate below. No pricing has been published; confirm on official pages before you commit. No affiliate links in this article.

The difference between a video generator and a world model is the difference between watching a film and playing a game. A video model hands you a finished clip. Genie 4 hands you a place: press a key and the camera moves, type a sentence and the world changes. That is why this launch matters even though almost nobody can use it yet — it is the clearest public sign of where interactive AI is heading.

What Genie 4 is

Genie 4 is Google DeepMind's fourth-generation world model, announced October 27, 2026, in a post authored by Raia Hadsell, Diego Rivas and Hang Qi. The lineage is short but fast: the original Genie (2024) generated tiny 2D platform games from a single image, Genie 2 added 3D-style worlds you could explore, and Genie 3 (2025) brought photorealistic, navigable scenes — but silent, and limited to about 60 seconds of coherence. Genie 4 is the first in the line with sound, live steering, and roughly double the runtime.

Unlike a traditional game, nothing in a Genie world is pre-built. There are no 3D models, no hand-coded physics, no pre-rendered assets. The model synthesizes each frame in real time as you navigate, building the environment directly from your prompt. DeepMind frames world models as a step toward AGI: understanding and simulating how the world behaves, not just describing it, is what lets robots and agents train in infinitely variable simulations where nothing breaks.

What is actually new

Native audio, generated with the world. This is the headline change and DeepMind's first for a real-time world model. Music, ambient noise and interactive dialogue are generated simultaneously with the visuals — not dubbed on afterwards. Characters in a scene can converse and react in real time, which turns a world from something you watch into something you can talk to.
Real-time steering with plain language. You can change the world mid-simulation with natural-language prompts: change the weather, move the camera, add a person or an object, set an event in motion — while the scene keeps running. Gemini adapts your prompts to fit the scene, so you steer in sentences, not in code.
Longer, more coherent scenes. Reported scene coherence is up to about two minutes, roughly double Genie 3's 60-second limit. Two minutes is still short, but it is long enough for a scene to develop — a character can cross a room, weather can turn, a conversation can go somewhere. The jump matters because drift has always been the failure mode of world models: the longer a simulation runs, the more details slip.
More realistic visuals. DeepMind describes enhanced photorealism and more fluid movement, with Gemini Omni anchoring the real-world appearance and behaviour of objects so they look and act more like the real thing.
SynthID watermarking. All generated video and audio is marked with Google's SynthID, so Genie 4 output carries a machine-readable watermark identifying it as AI-generated.

How it works under the hood

Two Gemini systems do the steering. Gemini Omni anchors simulations in real-world appearance and behaviour — it is why objects look and move the way they should. Gemini 3.8 Flash acts as the guiding agent inside the loop: it breaks down high-level instructions, plans multi-step events, and evaluates each output frame to keep the simulation on track as you move through changing scenes. In practice, that means your one-sentence prompt gets expanded into the model's internal plan for what happens next, and every frame is checked against that plan.

That architecture also explains the limits. A world model is only as stable as its frame-by-frame memory: every new frame is generated from what came before, so small errors compound. DeepMind is candid about this — it is the reason scenes top out around two minutes rather than running indefinitely.

How to try it (and who can't yet)

Trusted testers: Genie 4 is available now as a research preview to trusted testers only.
Google AI Ultra subscribers: coming soon through Project Genie, the consumer-facing product that launched in January 2026.
Developers: coming soon to early-access partners of the Gemini API and Google AI Studio.
Everyone else: a waitlist is open. No general-availability date has been announced.
Pricing: none published. Resolution, frame rate and API rates were not in the announcement either — treat any figures circulating online as unverified until Google publishes them.

What you could actually use it for

For creators, the near-term uses are exploratory rather than production-ready: rapid interactive storyboarding (walk through a scene instead of storyboarding it frame by frame), interactive demos and prototypes, and training environments for agents and robots. For game developers, Genie 4 is a prototyping tool, not a replacement — a generated world is not a shippable game, and the two-minute coherence ceiling rules out anything longer-form today. The bigger customer, as DeepMind itself signals, may be robotics: Gemini directs, Genie simulates, and a robot can practise in a world where mistakes cost nothing.

The honest caveats

We haven't tested it. Genie 4 is a closed research preview. Everything here comes from DeepMind's announcement and launch reporting, not our own runs.
Drift is still the limit. DeepMind itself names it: visual details and object permanence can shift or drift over time, especially in longer runs, after multiple complex prompt edits, or while navigating.
Named limits, in DeepMind's own words: complex physics, audio accuracy, legible text and spatial accuracy are all still imperfect.
Two minutes is a ceiling, not a floor. Coherence up to about two minutes is the reported best case; many sessions will degrade sooner.
No pricing, no specs. Resolution, frame rate, countries and API pricing were not published at launch. Anyone quoting them is guessing.
Not a game engine. A generated world has no persistent state, no save files, no hand-tuned gameplay. It is a simulation you visit, not a product you ship.

Verdict

Genie 4 is the most interesting AI launch of the week — and the least usable. Native audio plus live steering plus two-minute coherence is a genuine step: for the first time, a world model feels less like a tech demo and more like a place. But it is a closed research preview with no published pricing, no general-availability date, and drift limits DeepMind openly admits. The honest move: join the waitlist if you are curious, watch what trusted testers publish, and do not plan any production work around it until Google opens access and independent testers can run it themselves.

FAQ

What is Google Genie 4?

A world model from Google DeepMind, announced October 27, 2026. It generates interactive, photorealistic virtual worlds in real time from a text or image prompt, with native audio — music, ambient sound and spoken dialogue — generated alongside the visuals. You navigate with keyboard controls and can change the world mid-simulation with new text prompts.

Can I try Genie 4 today?

Only if you are a trusted tester. Genie 4 launched as a research preview for trusted testers, and 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. A waitlist is open for everyone else.

How is Genie 4 different from Genie 3?

Three upgrades: native audio (Genie 3 was silent), real-time steering with natural-language prompts mid-simulation, and scenes that stay coherent for up to about two minutes — roughly double Genie 3's 60-second limit. Visuals are also more photorealistic, with smoother movement.

How much does Genie 4 cost?

No pricing has been published. It is a research preview for trusted testers, with wider access coming through Google AI Ultra (Project Genie) and the Gemini API / AI Studio for early-access partners. Until Google publishes API rates or plan details, treat any price figure you see elsewhere as speculation.

Is Genie 4 a game engine?

No. A game engine runs hand-built 3D assets and hand-coded rules. Genie 4 synthesizes the world frame by frame as you move through it — nothing is pre-built. That is also why it drifts: details and object permanence can shift over longer sessions, and complex physics, legible text and spatial accuracy are named limits by DeepMind itself.