Google DeepMind's Genie 4 adds native audio and real-time steering to generated worlds

Real-time worlds from a prompt

Google DeepMind has announced Genie 4, a world model that produces explorable, photorealistic environments in real time, starting from a text prompt or a single image. According to a dev.to writeup of the October 27, 2026 announcement, the model is available as a research preview for trusted testers, rolling out first to Google AI Ultra subscribers through Project Genie and to early-access partners of the Gemini API and Google AI Studio, with a waitlist open to everyone else.

Native audio and longer scenes

Two changes define the release. The first is sound: Genie 4 is DeepMind's first real-time world model with native audio, meaning music, ambient sound and interactive dialogue are produced alongside the visuals rather than layered on afterwards. Characters in a scene can converse and react as the simulation runs.

The second is duration. Coherent scenes now hold together for up to about two minutes, twice the roughly 60-second limit of Genie 3. The dev.to report frames this as significant because drift has always been the weak point of world models — the longer a simulation runs, the more details slip. The writeup notes DeepMind is upfront that the problem is not solved: visual details and object permanence can still shift during longer runs, after several complex prompt edits, or while the user is moving through the scene.

Steering while the simulation runs

The model can also be directed mid-run. Users can alter the weather, reposition the camera, or introduce new people and objects without pausing the scene, issuing instructions in natural language while navigating with keyboard controls. Gemini adapts those prompts to fit the scene, according to the writeup. On the architecture side, Gemini Omni anchors how real-world objects look and behave, while Gemini 3.8 Flash acts as an in-loop guiding agent that keeps the simulation on track.

Better people, physics and provenance

DeepMind points to people and physics as the biggest advance in this generation: characters move more naturally, and the model simulates collisions and aerodynamics more convincingly than before. Every video and audio output from Genie 4 carries SynthID watermarking, so generated material is identifiable as AI-made — a detail the dev.to report argues will matter more as these tools close in on photorealism.

Why it matters

The near-term applications are suggestive rather than confirmed, and the dev.to writeup is explicit about that distinction. In game design, a director could sketch a playable scene from a single sentence and iterate on it live instead of waiting for a build. In film production, teams could block out locations, test camera moves and rehearse sequences before any set exists. In education, students could walk through a historical street or around a physics experiment.

Robotics is the one area with a concrete research link. DeepMind paired Genie 4 with a Gemini 4 Argon model that breaks tasks into steps simulated in real time, a possible route to testing robot behaviours in simulation before any hardware is involved.

The honest assessment is that Genie 4 remains a research preview with firm limits: two minutes of coherence, drift under pressure and a restricted rollout. But the direction of travel is clear. Worlds that look real, sound real, respond to conversation and accept live direction are edging away from lab demonstrations and toward becoming a practical prototyping medium — and the useful skill may turn out to be learning how to direct them while the coherence window is still measured in minutes.