The Pulse
Google DeepMind Unveils Genie 4 With Real-Time Audio
Google DeepMind announced Genie 4 on October 27, 2026, describing a world-simulation model that generates interactive scenes, ambient sound and dialogue. The research preview is limited to trusted testers, while Google says access through P
AI.info Team · Written with AI · fact-checked ·
Google DeepMind announced Genie 4 on October 27, presenting it as a real-time world model that generates interactive scenes along with their sound. Users can steer a simulation with text prompts and keyboard controls, while the system produces ambient audio and dialogue as events unfold.
Text prompts steer worlds as they run
Genie 4 lets users change a scene in progress by asking for a shift in weather, a camera move or a new person or object. Google says Gemini adapts those instructions to the surrounding scene and monitors the simulation between user inputs to help it progress coherently. The announcement describes users navigating and controlling the camera with standard keyboard commands.
Google DeepMind says Genie 4 builds on the Gemini family: Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models contribute reasoning intended to preserve coherence and continuity. In a more specific setup, Gemini 3.8 Flash acts as a guiding agent, observing frames, breaking goals into steps and checking whether each step succeeds. The company says that agent can revise its plan when the scene changes or a user intervenes.
Native audio expands the simulation
DeepMind describes Genie 4 as its first real-time world model to generate audio natively. That includes environmental sounds, such as wind or waves, as well as character dialogue that can respond to user input and events. The company says the sounds accompanying its demonstration videos were generated directly by the model.
The post also describes experiments combining Genie 4 with Gemini 4 Argon and Gemini Robotics. In one example, Argon breaks a task into steps for Genie 4 to render as a simulation of Apptronik’s Apollo humanoid playing checkers. DeepMind presents the work as research into how AI systems might test actions in simulated settings before they are attempted in the physical world.
Preview access comes before wider release
Genie 4 is rolling out first as a research preview for trusted testers. Google says access is coming soon to Google AI Ultra subscribers through Project Genie, along with early-access partners using the Gemini API and Google AI Studio. The announcement offers a waitlist for Project Genie but does not give a date for broader access.
DeepMind says it evaluated the model internally for harmful or malicious content and is starting with trusted testers to gather feedback. Generated video and audio will include SynthID, Google’s watermarking technology for identifying AI-generated media. The company does not publish model size, latency figures or benchmark results in the announcement.
DeepMind flags drift and weak long-term consistency
The company cautions that Genie 4 is experimental. Environments can remain coherent during short sessions, but visual details and object permanence may drift as sessions lengthen, prompts become complex or users move through a scene. DeepMind also lists complex physical interactions, accurate real-time audio, legible text, spatial accuracy and long-term consistency as active research problems.
While it can maintain coherent environments over short sessions, visual details and object permanence can still shift or drift over time—especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene.
Raia Hadsell, Diego Rivas and Hang Qi
Those limitations matter for the uses Google highlights, from interactive creative previews to robotics and agent training: a simulation that changes details or loses consistency over time may not reliably reproduce a scenario. For now, DeepMind is offering Genie 4 to trusted testers, with the broader user and developer preview still to come.