On October 27, 2026, Google DeepMind announced Genie 4, a new world model that generates explorable, interactive worlds in real time — and, for the first time in the Genie line, generates the sound to go with them. Type a prompt, step into the scene with keyboard controls, and the world responds: characters move, weather changes, dialogue adapts, and music and ambient noise are produced as part of the same simulation.

There is a catch, and it is an important one. Genie 4 is a research preview, rolling out first to trusted testers, with Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio next. Everyone else can join a waitlist. The smart move for businesses is therefore not to plan a product around Genie 4 today, but to understand what it actually does, where it breaks, and how to evaluate it before you build anything on it.

What is a world model?

A world model is not a video generator, and it is not a game engine. A video generator produces a clip you watch from start to finish. A game engine renders a world a designer built in advance, from hand-made assets and hand-written rules. A world model does something in between: it predicts, frame by frame, what a place should look and sound like as you move through it and act in it, generating the environment on the fly from a text prompt or an image.

That distinction explains both the excitement and the limits. Because nothing is pre-built, you can go from an idea to a prototype in seconds. Because everything is predicted, the world can drift: a door can change colour, an object can disappear, a character can forget where it started.

What is actually new in Genie 4?

1. Native audio: worlds that sound as well as look

DeepMind describes Genie 4 as its first real-time world model to generate audio natively. The sound is not a soundtrack added afterwards. Wind in the trees, waves on a beach, music, and spoken dialogue are generated simultaneously with the world itself, so a character’s words and the ambient sound of a place respond to what is happening in the scene.

A world you can hear is a world you can converse with: characters react to your inputs and to events in the scene in real time. Audio errors are also immediately obvious, and DeepMind lists real-time audio accuracy as an area of ongoing research, so expect rough edges here first.

2. Live steering: rewrite the scene while it runs

Genie 4 lets you modify a running simulation with natural language. Change the weather, move the camera, introduce a new person or object, and Gemini adapts your prompt to fit the surrounding scene. Between your inputs, Gemini also monitors and steers the scene so the world keeps progressing coherently rather than stalling between prompts.

This has the most immediate practical value for creative teams. A director can run multiple takes, adjusting camera, lighting, and dialogue on the fly; a game designer can test a style or mechanic before writing engine code; a product team can show a stakeholder three versions of a concept in one meeting. When each iteration costs seconds, teams can afford to try far more ideas.

3. More realistic physics and people

DeepMind says Genie 4 simulates real-world physics scenarios such as aerodynamics and collisions, and calls its simulated people the most dramatic leap in this release. Characters move naturally, respond organically to their environments, and can be steered in real time.

People are the hardest test for any world model: a human walking and speaking is judged instantly. If Genie 4’s people really move more naturally, that matters for robot training, service simulations, and education too. Treat the claim as promising but unproven at scale — DeepMind’s demos are curated, and complex physics remains an area of ongoing research.

4. Gemini inside the loop: reasoning, not just rendering

The most technically interesting change is architectural. Genie 4 does not run alone. Gemini Omni anchors simulations in real-world appearance and behaviour, while other Gemini models supply the reasoning that keeps a scene coherent. In real time, Gemini 3.8 Flash serves as a guiding agent in the loop: it observes the scene, breaks a goal down into steps, reviews incoming frames to check whether each step succeeded, and adapts its plan when the world shifts or the user intervenes.

DeepMind is also experimenting with pairing Genie 4 with Gemini 4 Argon for robotics. In the published example, Argon breaks a task into steps fed to Genie 4 as captions while the simulation renders — Apptronik’s Apollo humanoid playing checkers. The idea is to test multi-step tasks in simulation before committing hardware, floor time, and safety risk to the real world: cheap, infinitely variable rehearsal spaces for agents and robots.

Capability What Genie 4 adds Why it matters
Native audio Music, ambient sound, and dialogue generated with the world Scenes become conversational, not just watchable
Live steering Natural-language edits mid-simulation, adapted to the scene Ideas can be tested in seconds, not review cycles
Realism Improved physics, collisions, and natural human movement Simulations become credible for training and pre-visualisation
Reasoning in the loop Gemini 3.8 Flash guides steps and checks progress frame by frame Multi-step tasks stay on track instead of drifting

World model, video generator, or game engine?

It helps to place Genie 4 against tools teams already use, because it replaces none of them outright. A video generator produces polished clips you watch; you cannot walk around inside them. A game engine produces repeatable, production-quality experiences, but every asset and rule must be built by hand first. Genie 4 sits between them: fast, explorable drafts and scenario tests, where a short, imperfect simulation is more useful than no simulation at all.

DeepMind’s trusted testers describe it that way: filmmakers use it for pre-visualisation, game developers to test a concept before writing engine code, and enterprise partners for rapid zero-to-one scenario generation. The world model shortens the path to a decision; a traditional tool still produces the final product. Think draft mode, not delivery mode.

Who can use Genie 4, and when?

Access is narrow for now, and that is deliberate. DeepMind has started rolling Genie 4 out as a research preview for trusted testers, gathering feedback before a broader release. Next are Google AI Ultra subscribers in Project Genie and early-access partners of the Gemini API and Google AI Studio. A waitlist for Project Genie is open.

Three details matter for planning. First, DeepMind has not published pricing, benchmarks, resolution, or frame-rate specifications, so any number you see for those is from a report or an earlier model, not DeepMind. Second, the reported coherence window — about two minutes, roughly double Genie 3’s reported one minute — comes from independent coverage; DeepMind itself says only that Genie 4 maintains coherent environments over short sessions. Treat two minutes as a rule of thumb, not a specification. Third, all Genie 4 video and audio will include SynthID, Google’s imperceptible digital watermark.

What businesses can actually do with it now

If you get access, the highest-value uses are the ones where a short, imperfect simulation still beats no simulation at all.

  • Pre-visualisation. Step into a live scene, run multiple takes, and settle camera, lighting, and dialogue before a shoot — a decision aid, not a final frame.
  • Game prototyping. Turn an idea into an explorable prototype in seconds and only commit engineering time to concepts that survive contact with players.
  • Robot and agent training. Generate varied environments for multi-step tasks and test how an agent behaves when the world changes mid-task — cheaper and safer than real trials, if validated on real hardware.
  • Education and history. DeepMind notes photojournalists exploring how one archival photograph can become a grounded simulation, or how a flash flood might unfold — tangible, if clearly labelled.
  • Enterprise scenarios. Sketch a warehouse or emergency scenario, walk through it, and find problems a floor plan hides — surfacing questions early, not producing a certified safety model.

For most small businesses, the honest near-term use is simpler: use the waitlist period to build a prototyping habit. Teams that already experiment with AI utilities will adapt faster than teams waiting for a finished product. Our guide to free AI tools, our WebReOn Tools collection, and WebReOn Chat are useful starting points while you wait.

What Genie 4 cannot do yet

DeepMind’s limitations section is short and worth taking seriously: 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 a scene. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are all active areas of ongoing research.

In business terms, long scenes will not stay consistent, text inside a world may not be readable, layouts may not match a real floor plan, and sound may be plausible rather than accurate. A convincing simulation is not evidence that a plan will work in the real world. Genie 4 is therefore a tool for exploration, not certification, compliance, or safety-critical decisions — and building a deliverable on a preview with no published price or firm release date means building on a moving target. Prototype with it, document what you learn, and keep production work on tools you control.

A practical checklist for your first session

  1. Pick one question, not one demo. Decide what you need to learn: does this camera angle work, does this layout confuse people, does this agent recover when the scene changes?
  2. Run the same prompt three times. World models are not deterministic. Variation between runs tells you how much to trust any single result.
  3. Test the failure modes on purpose. Push past two minutes, make several edits in a row, and put readable text in the scene. Note exactly where it breaks.
  4. Check the audio separately. Listen with your eyes closed first. Does the dialogue make sense? Does the ambient sound match the place?
  5. Label everything you share. SynthID watermarks the output, but add your own clear label too. Anyone who sees a simulation should know it is AI-generated and provisional.
  6. Validate before you act. If a simulation suggests a layout or a robot behaviour, confirm it in the real world before spending money or changing a process on its strength.

What to watch next

Three milestones will tell you whether Genie 4 becomes a working tool or stays a research demo: the breadth, cost, and limits of the Ultra and API rollout; the hybrid world model direction DeepMind sketches for the future, combining structured reasoning with pixel-level simulation, which could extend coherence well beyond a couple of minutes; and evidence from robotics that simulated training measurably improves real robot performance.

Until then, be curious but disciplined. Genie 4 shows interactive worlds with sound, speech, and reasoning are no longer science fiction — but not yet reliable enough to run a business on. The teams that benefit first will learn fast, share honestly labelled prototypes, and keep production decisions grounded in the real world.

Frequently asked questions

Is Genie 4 available to the public now?

No. As of the October 27, 2026 announcement, Genie 4 is a research preview for trusted testers. Google AI Ultra subscribers in Project Genie and early-access Gemini API and Google AI Studio partners are next, and a waitlist is open.

How long can a Genie 4 world stay consistent?

DeepMind says Genie 4 maintains coherent environments over short sessions and warns that details can drift over longer runs. Independent coverage reports about two minutes, roughly double Genie 3’s reported one minute, but DeepMind has not published that figure as a specification, so plan demos conservatively.

Can Genie 4 replace a game engine or a video generator?

Not yet. A game engine still produces repeatable, production-quality experiences, and a video generator still produces polished clips. Genie 4’s strength is fast, explorable drafts and scenario testing.

Is Genie 4 output labelled as AI-generated?

Yes. DeepMind says all video and audio generated by Genie 4 will include SynthID, its imperceptible digital watermark. You should still add your own clear label when sharing simulations.

Should a small business invest in world models today?

Invest time, not dependence. Join the waitlist, prototype one real question if you get access, and keep production work on tools you control until pricing, limits, and long-session reliability are published and proven.

Sources

Conclusion

Genie 4 is best understood as a threshold, not a product launch. A world you can step into, talk to, and rewrite while it runs — with sound generated alongside the picture and a reasoning model steering the scene — is a new kind of tool. Its limits are equally clear: short sessions, drifting details, unreadable text, and no published price or specification. For businesses, the strategy is clear: experiment early, evaluate honestly, and build on what you learn rather than on the preview itself.