Google DeepMind's Genie 4 adds sound, speech and a Gemini planner to real-time world simulation

Google DeepMind announced Genie 4 on October 27, its new world model: a system that generates an explorable environment frame by frame, in real time, while the user moves through it with the keyboard and changes it with text prompts. The release is a research preview. Only a group of trusted testers has access today; Google says the model is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, without giving a date. A waitlist is open.

What is new is less the idea of a playable generated world, which Genie 3 already showed in August 2025, than what is layered on top: sound, speech, more realistic people, and a Gemini model sitting in the loop to keep the simulation coherent.

What Google says Genie 4 adds

• Native audio. Google describes Genie 4 as its first real-time world model to generate audio itself: ambient sound for each setting, music, and dialogue. Characters can talk to each other and react to the user or to events in the scene.
• More realistic people and physics. DeepMind calls the way Genie 4 renders people its biggest leap, with characters that move naturally and can be steered live. It also cites physical scenarios such as collisions and aerodynamics.
• Live steering in plain language. Mid-simulation, a user can type an instruction such as changing the weather, moving the camera or adding a person or object. Google says Gemini adapts that prompt to the surrounding scene before it reaches the simulator.
• Longer consistency, per TechCrunch. TechCrunch reports that Genie 4 keeps a scene consistent for up to two minutes, double Genie 3. Google's own posts do not state a duration.

All of these are Google's descriptions, backed by demo videos Google selected. There is no published benchmark, model size, resolution or frame rate, and no independent evaluation yet.

A simulator with a planner on top

The architectural detail worth attention is the split of work. In the DeepMind post, Gemini 3.8 Flash acts as a guiding agent: it watches the scene as Genie 4 renders it, breaks a goal into steps, checks incoming frames to verify each step succeeded, and revises the plan if the world changes or the user intervenes. Genie handles pixels and sound; Gemini handles structure and narrative.

DeepMind pushes the same pattern toward robotics. In one research demo, Gemini 4 Argon breaks a checkers task into steps and feeds them to Genie 4 as text captions, and Genie 4 simulates Apptronik's Apollo humanoid playing the game. The Decoder notes that no physical robot is involved; it is a simulation of one. DeepMind presents the combination as an early step toward hybrid models that pair language and code reasoning with pixel-level simulation, and as a way to test how AI systems handle an environment before they meet the real one.

For anyone building agents, the loop will look familiar: plan, act, observe, verify, re-plan. The difference is that here the environment itself is generated, so a training or test world can, in principle, be varied without limit. Whether those worlds are physically faithful enough to be useful for robot training is exactly what has not been shown publicly yet.

What Google admits does not work yet

DeepMind's post is unusually direct about limits. It calls Genie 4 an early, experimental model that stays coherent over short sessions, but says visual details and object permanence can drift, especially in longer runs, after several complex prompt edits, or while moving through the scene. It lists complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency as open research problems. TechCrunch adds that Google engineer Diego Rivas stressed in a briefing that the end-to-end product experience is not yet clear.

On safety, Google says Genie 4 went through internal safety evaluations, ships first to trusted testers deliberately, and that all generated video and audio will carry its SynthID watermark.

Who it is for, for now

Google and its testers point to pre-visualization in film, turning an archival photograph into an explorable scene, prototyping game mechanics before writing engine code, and enterprise scenario simulation. TechCrunch also places the launch in a busy field: Runway shipped a playable world model with audio a couple of months ago, and AMD agreed in September to buy World Labs.

What changes for developers

Not much today, since there is no public API access. Two things are worth tracking. First, Gemini API access is promised to early access partners, which would make world simulation something a developer can call rather than only watch in demos. Second, prompting here is continuous rather than one-shot: instructions arrive while the world is running and are rewritten by another model to fit the scene. That is closer to directing a live agent than to writing a single video prompt, and it is a pattern likely to spread beyond Genie.