Imagine walking through a scene while it is being generated, changing its weather and asking a character to act without starting over. That is the experience Google DeepMind is presenting with Genie 4, announced on October 27. The model is entering a research preview for selected testers, adding native audio and natural-language changes during an interaction.
The appeal extends beyond attractive imagery. A creative team could use an explorable setting to discuss camera placement, movement and sequence before committing to production. Agent researchers could investigate varied situations in a generated environment. These remain exploratory uses, and their value depends on how consistently that environment behaves.
A world that responds as you explore
World models attempt to represent how an environment evolves in response to actions. Genie 3 already supported interactive exploration, so navigation itself is not new. Genie 4's proposed advance brings together sound, dialogue and more control over a running simulation.
On the model page, Google shows prompts changing landscapes, cameras, weather and characters. Sound is generated alongside the environment rather than simply added as a soundtrack afterward. That changes what a rough draft can convey: a scene can be considered through its action and its sound.
The system also divides work between models. DeepMind describes Gemini observing frames, breaking goals into steps and adapting its plan as the scene changes. That coordination helps explain the approach: generating an appearance and organizing a sequence of actions require different capabilities. The published examples are still company demonstrations.
Visual realism is only one kind of fidelity
A convincing simulation needs to be judged against its intended use. Some instability may be acceptable when choosing a camera angle. Establishing that a robot can manipulate an object requires much closer scrutiny of materials, distances and the consequences of each movement.
Google acknowledges that details and objects can drift during longer sessions or after complex interventions. Spatial accuracy, physical interactions, audio and consistency remain under development. An action looking plausible on screen therefore provides limited evidence about whether it would work outside that environment.
A useful next test would repeat a sequence, change one condition and check whether the consequences remain consistent. An object disappearing or changing when the camera returns could undermine a story as well as an agent evaluation. That less spectacular form of continuity is where the tool will need to earn confidence.
Who can try it
Current access remains restricted. DeepMind plans to expand to Google AI Ultra subscribers through Project Genie and early-access partners in the Gemini API and Google AI Studio. The official model page provides a waitlist; joining it does not mean the model is available to every subscriber.
Creative teams can watch whether experimentation makes it easier to test an idea and discuss alternatives. Engineering teams face a further question: how should the environment's fidelity be measured before it is used to evaluate decisions? Genie 4's value will depend on making those differences observable so each project can establish what it can rely on.
Sources: Genie 4 announcement, model page and Genie 3 introduction.