On October 27, Google DeepMind introduced Genie 4, a model that creates interactive scenes with sound and lets people change them while the simulation is running. Google's examples show users moving through a world, changing the weather with a text prompt and watching characters move and speak. For now, the model is available only to a limited group of testers.
An official Google DeepMind demonstration: a virtual scene with sound and camera controls.
What interests me most is the chance to intervene while a scene is still unfolding. Google's polished examples show the direction of the research, but we will need wider access before judging how practical the model is for everyday work.
What Genie 4 can do
Picture a path through a forest. You can use a keyboard to move and look around, then ask the model to change the weather or add an object. Google shows text prompts changing a scene as it runs: the generated world continues to develop with the new instruction in mind.
Alongside the visuals, Genie 4 generates ambient sounds, music and dialogue. In the published demonstrations, characters can speak and react to events. That is part of the model's design: sound is generated as part of the scene rather than necessarily being a separate track added after generation.
Google calls Genie 4 a world model. These systems generate an environment and attempt to show how it changes in response to user actions or other events. Genie 3 already supported interactive steering; the fourth version adds native audio generation. Google has collected the details and examples in its Genie 4 announcement.
How an interactive world differs from a video
A finished video has a fixed sequence of frames. You can replay it after it has been made, but pressing a key will not change what happens inside it. In Google's Genie 4 demonstration, the scene is generated during interaction: the user sets a direction or provides another instruction, and the model continues the simulation in response.
Models from the Gemini family also help maintain the flow of events. Google says Gemini helps interpret instructions in the context of the current scene and plan what happens next. That does not mean every item in the world behaves like a separate object in a conventional game engine; the practical distinction is that the scene can respond to a user's input.
Suppose you need rain instead of sunshine for a storyboard. With a conventional video generator, you would usually prepare a new prompt and generate another clip. Genie 4 is intended to let a user make that kind of change while exploring the scene. For now, Google's demonstrations are the evidence available for judging how the approach works.
Where the model still struggles
The virtual world does not always preserve details over longer interactions. Google says objects and visual details can shift or drift, especially as a session runs longer or after several complex instructions. Accurate physics, spatial relationships, sound and readable text are also areas that need more work.
That matters if a project depends on a particular object staying in place or a sequence of actions happening correctly. A convincing movement in a demonstration does not establish that the model will handle every complex scenario with the same accuracy. Any result with precise requirements will need careful review.
Who can use Genie 4?
At the time of writing, Genie 4 is a research preview for trusted testers. Google plans to expand access through Project Genie to Google AI Ultra subscribers and to early access partners for the Gemini API and Google AI Studio. Its announcement does not give a precise expansion date or a separate price for Genie 4.
The announcement page has a waitlist. Having access to other Gemini models or earlier Project Genie versions does not by itself mean that a particular user can already use Genie 4. We did not conduct our own Genie 4 tests for this article.
What this means for you
For people developing videos or games, this kind of interaction could help during early ideation: explore an environment, change the weather and try another camera angle before preparing the final material. Google also sees simulations as a research environment for AI agents and robots. Those are research directions rather than tools already available to everyone.
If you need a finished clip today rather than an interactive simulation, explore the available video models or our guide to making AI video. The workflow is different: you generate a clip from a prompt and then assess the result.
Sources
- Google DeepMind: Genie 4: A new generation of world simulation - the official announcement, including capabilities, limitations and access; October 27, 2026.
- Google DeepMind: Genie 3 - the previous model's capabilities, including controls and promptable changes to the world.