Google DeepMind Unveils Genie 4, a World Model You Can Steer and Hear
Google DeepMind announced Genie 4 on October 27, 2026, a new world model that generates interactive scenes in real time and, for the first time in the Genie line, generates the sound to go with them. Instead of producing a clip to watch, Genie 4 produces a place to move through: users navigate with standard keyboard controls and change the scene with text prompts while it keeps running.
The announcement matters because it moves world models closer to practical tools. A video generator shows what a scene looks like. A world model tries to answer what happens next when someone acts inside it. Genie 4 adds sound, live steering, and a Gemini model inside the loop to keep multi-step events on track, while remaining an early research preview rather than a finished product.
What Genie 4 is
DeepMind describes Genie 4 as a world model for real-time interactive simulation. A user sets an event in motion, and the model simulates how the world responds, how it sounds, and how the people in it move, react, and speak. Music, ambient noises, and interactive dialogue are generated at the same time as the world itself, not added afterwards.
That is the headline change from earlier Genie models. DeepMind says Genie 4 is its first real-time world model to generate audio natively, and characters converse and react to inputs or events in the world in real time.
Four capabilities in the announcement
DeepMind groups the advance into four areas. First, real-time steering: users can modify a simulation with natural language, for example by changing the weather, moving the camera, or introducing a new person or object. Gemini adapts the prompt to fit the scene and keeps the world progressing between inputs.
Second, realism for physics and human action. DeepMind says Genie 4 simulates scenarios such as aerodynamics and collisions, and that the most dramatic leap is in simulated people, who move naturally and can be steered in real time.
Third, native audio: ambient sound, music, and dialogue that adapts on the fly.
Fourth, complex events simulated with Gemini models. Gemini Omni anchors the simulations in real-world appearance and behavior, while other Gemini models help maintain coherence. Gemini 3.8 Flash serves as a guiding agent in the loop: it observes the scene, breaks goals into steps, and reviews frames to verify that each step succeeded, adapting its plan if the world shifts or the user intervenes.
DeepMind also describes an experiment pairing Gemini 4 Argon with Genie 4, in which Argon breaks a task into steps for Genie 4 to render. The example is Apptronik’s Apollo humanoid playing checkers, with steps interleaved as text captions while the simulation renders.
How long a world holds together
One number to treat with care is scene length. DeepMind says Genie 4 can maintain coherent environments over short sessions, but it does not present a single session length as an official specification in the announcement reviewed here. Reporting by TechCrunch and The Verge, as summarized by Skimwire, puts consistent scenes at up to about two minutes, about double the roughly one minute reported for Genie 3. That is a reported press figure, not an independently tested result, since no outside tester has had general access yet.
Who can try it, and what it cannot do yet
Genie 4 is available now only as a research preview for trusted testers. DeepMind says it is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, and that a waitlist for Project Genie is open. No firm general availability timeline was announced in the material reviewed.
DeepMind is explicit about the limits. Visual details and object permanence can still shift or drift over time, especially in longer runs or after multiple complex prompt edits. Complex physical interactions, real-time audio accuracy, legible text, spatial accuracy, and long-term consistency remain active areas of research.
On responsible use, DeepMind says Genie 4 includes safety guardrails consistent with its policies for other models, and was evaluated through internal safety evaluations before the preview. All video and audio generated by Genie 4 will include SynthID, Google’s imperceptible watermark for identifying AI-generated content.
Why it matters
For filmmakers, Genie 4 is being tested as a rapid draft mode for pre-visualization: step into a live simulation and run multiple takes in real time, adjusting camera angles, lighting, dialogue, and action. Game developers are testing it to go from an idea to an explorable prototype in seconds, before writing engine code.
The larger bet, in DeepMind’s framing, is simulation for robots and agents. If a model can generate variable training environments, teams can test how AI systems handle multi-step tasks before bringing them into the real world.
None of that makes Genie 4 a replacement for a game engine, a film crew, or a physics simulator today. Two minutes of coherence, drift over longer runs, and a preview-only rollout mean the sensible use right now is prototyping and testing ideas, not shipping finished experiences. The release is still a clear signal: the next step after chatbots and video clips may be simulated worlds that users can redirect mid-scene with plain language.
