What changes compared with the past
Google DeepMind has presented Genie 4, successor to Genie 3 shown in August 2025. According to The Decoder, it is not a generator of finished videos: the model computes the scene frame by frame and reacts to inputs in real time. The user navigates the environment with the keyboard and steers events via text instructions.
The main new feature cited by the company is directly generated audio: each world would have its own background sound, such as wind among the trees. Genie 4 also handles dialogue, with characters talking to one another and reacting in real time to inputs or simulation events. DeepMind also claims a leap in the realism of how people are rendered; they should move naturally and be directly controllable.
Gemini's role behind the scenes
Its operation relies on other Gemini models. Genie 4 leverages capabilities of the Gemini Omni video model for the world's appearance and behavior, while other models plan the simulation to keep it coherent. If a user changes the weather with a prompt, Gemini would adapt the instruction to the current scene and guide events between one input and the next.
The example provided concerns multi-step tasks: Gemini 3.8 Flash observes the scene, breaks the objective into stages and checks on the new frames whether each one has worked, recalculating the plan if the world changes or the user intervenes. DeepMind is also experimenting with Gemini 4 Argon to connect work on world models to Gemini Robotics robotics research: in a demo, Argon breaks a game of checkers into text instructions that Genie 4 uses to simulate Apptronik's Apollo humanoid robot, with no physical robot involved.
Stated limitations and availability
DeepMind describes Genie 4 as an experimental and early-stage model. Coherence holds in short sessions, but in longer simulations or after several complex prompt changes, visual details may change and objects do not always persist. The company says it is still working on complex physical interactions, accurate real-time audio and legible text.
For now, access is reserved for selected testers as a research preview. It is expected to arrive for Google AI Ultra subscribers in the Project Genie web app, together with partners with early access to the Gemini API and Google AI Studio. Google assures that videos and audio tracks will carry the invisible SynthID watermark. It remains to be verified in the field how well the claims about coherence and realism hold up in prolonged use; the material does not offer independent measurements.
