Google DeepMind announced Genie 4 on October 27, 2026, and here is the sentence that actually matters: the world now comes with sound. For the first time in DeepMind's real-time world-model line, the simulation and the audio are generated together — the ambient noise of a room, music, and characters you can talk to — instead of a silent video with sound bolted on afterwards.

That is a real step. It is also, for almost every developer reading this, a waitlist story rather than a tool you can open this week. I am going to keep both of those facts in view at the same time, because most launch coverage will only give you one of them.

Disclosure: no affiliate links in this article. I have not used Genie 4 — it is in research preview — and the capabilities below are Google DeepMind's claims from its October 27, 2026 announcement (authors Raia Hadsell, Diego Rivas and Hang Qi), not independently verified results.

What Genie 4 actually is

Genie 4 is a world model: you give it a starting point and it generates an interactive, simulated world in real time, one frame after another, that you can move through. DeepMind's earlier Genie models did the visuals. Genie 4 adds three things that change what the thing feels like to use.

  • Native audio, generated with the world. Ambient sounds, music and interactive dialogue come out of the same generation process as the visuals. If you have watched a Genie 3 clip, you know the eerie part was the silence. That is gone — or at least, it is gone in DeepMind's demos.
  • More convincing worlds. Enhanced photorealism, more fluid movement, and better physics — aerodynamics and collisions get named specifically. The biggest leap, by DeepMind's own telling, is simulated people: how they look, how they move, how they behave around you.
  • Live steering. While a scene is running, you can change it with natural-language prompts: move the camera, change the weather, bring in a new character. You are not just watching a generated clip; you are directing it while it plays.

Runtime is about 2 minutes per simulation, roughly double Genie 3's ~1 minute. Two minutes still is not a game session or a film. It is a scene. Keep that scale in mind whenever a demo makes it feel endless — the clip ends, and you start a new one.

How it works, in plain terms

Two Gemini pieces sit underneath Genie 4, and the division of labour is worth understanding because it explains both the strengths and the failure modes.

Gemini Omni anchors the simulation in how the real world looks and behaves — the grounding that stops the world dissolving into dream logic. Gemini 3.8 Flash then guides coherence frame by frame, keeping the scene consistent as it unfolds: the same character, the same room, the same weather you asked for, from one moment to the next.

That split is also where you should expect the rough edges. World models are famous for a specific kind of failure: everything looks right for ten seconds, then a door changes place, a character's face drifts, a thrown object forgets gravity. DeepMind's framing is an attempt to push that failure further out. Whether it succeeds on scenes you design, rather than scenes DeepMind chose for a launch video, is exactly what nobody outside the company can verify yet.

The honest availability picture

This is the section most launch coverage will soften, so here it is plainly:

  • Now: research preview, available to trusted testers only.
  • Coming soon: Google AI Ultra subscribers, inside something called Project Genie.
  • Coming soon: early-access partners of the Gemini API and Google AI Studio.
  • Open now: a waitlist for Project Genie.

That is the whole list. There is no public API endpoint today, no pricing, no general signup that gets you in this week. If you are a developer who wants to build on this, the realistic move is not "start building" — it is "get on the waitlist and watch the Gemini API / AI Studio early access." Anyone telling you otherwise is selling something.

One more detail worth knowing: all generated video and audio carries Google's SynthID watermark. That is good practice, and it also tells you something about how Google expects this to be used — generated worlds that are identifiable as generated, not passed off as footage.

Why developers should care anyway

A waitlist story can still be worth your attention if it changes what becomes possible. Genie 4 does, in one specific area: training environments.

Robots and agents learn by doing, and doing things in the real world is slow, expensive and occasionally destructive. A simulated world you can vary infinitely — different rooms, different weather, different people walking through — lets a robot practise thousands of variations of a task without ever touching a real object. DeepMind says it tested Genie 4 with Gemini Robotics and with Apptronik's Apollo humanoid. That is the use case to take seriously: not a demo of a robot in a fake living room, but the idea that the fake living room becomes the training gym.

The same logic reaches two other fields. Film previsualisation: a director blocks a scene, changes the weather mid-take, moves the camera, all before a crew is booked. Game prototyping: a designer walks through a level that does not exist yet, talks to a character who was not written yet, and decides in an afternoon whether the idea is worth building properly. Both are real, and both are still "try it when you can get in" rather than "ship it now."

What Genie 4 is not, yet: a production tool. Two-minute runtimes, preview-only access, and vendor-only testing mean the right response is curiosity with a calendar reminder, not a roadmap rewrite.

What to watch — and what to do this week

  • Join the Project Genie waitlist if you want a shot at the Ultra-subscriber rollout. It costs nothing and it is the only door actually open.
  • Watch the Gemini API and Google AI Studio early access — that is the door that matters for developers, because it is the one that turns a demo into something you can call from code.
  • Do not build a product plan on Genie 4's demo behaviour. No pricing, no rate limits, no API surface and no independent testing exist yet. A plan built on a launch video is a plan built on marketing.
  • When you do get in, test the failure modes first. Long scenes, crowded scenes, fast camera moves, characters who leave and come back. That is where world models break, and that is what will decide whether Genie 4 is a toy or a tool for your use case.
  • Keep an eye on the audio story specifically. Native dialogue in a generated world is the least-proven part of this launch — and if it holds up, it is the part that changes game prototyping and training most.

Sources: Google DeepMind — Genie 4: A new generation of world simulation · Google Blog — Introducing Genie 4