Genie 4: DeepMind's World Model Now Makes Sound, Takes Orders Mid-Scene, and Rehearses Robot Tasks

Google DeepMind released Genie 4 on Tuesday, 27 October 2026, as a research preview for a small group of trusted testers. The company calls it “our most realistic and interactive world model yet.” A world model generates an explorable environment frame by frame in real time and reacts to what the user does, unlike a video model, which renders a fixed clip. Genie 4 adds three things its predecessor lacked: sound, text control while the scene runs, and a Gemini agent that plans and checks what happens next.

What happened

The announcement came as a DeepMind blog post signed by Raia Hadsell, Diego Rivas and Hang Qi. TechCrunch and Unite.AI covered it the same day. Genie 3, the previous version, was announced in August 2025. It ran at 720p and 24 frames per second, kept a scene coherent for roughly a minute, and had no audio. Since 29 January 2026 it has been available to Google AI Ultra subscribers through a Labs site called Project Genie, which caps each exploration at 60 seconds.

Genie 4 is not publicly available yet. DeepMind says the model is “coming soon” to AI Ultra subscribers in Project Genie and to early-access partners of the Gemini API and Google AI Studio. TechCrunch puts the timing at “the coming weeks.” There is a waitlist, but no price and no date.

What is technically new

Native audio. DeepMind describes Genie 4 as its “first real-time world model to generate audio natively”. Music, ambient sound and character dialogue are generated at the same time as the frames, not added afterwards. A scene’s sound therefore comes from the same model that knows a door just slammed or a character just turned around. DeepMind also says it has not yet solved audio accuracy.

Steering while the scene runs. You navigate with the keyboard as before, but you can also type prompts while the simulation is running: change the weather, move the camera, bring in a new person or object. According to the blog, Gemini “will contextually adapt your prompt to match the surrounding scene”, so a bare instruction is rewritten to fit what is on screen. Genie 3 already had “promptable world events”, but Genie 4 turns them into free-form edits of a scene that is still running.

Longer coherence. TechCrunch reports that Genie 4 keeps a scene consistent for up to two minutes, double Genie 3’s limit. DeepMind’s own post makes a weaker claim, “coherent environments over short sessions”, and Unite.AI repeats that wording without a number. Treat two minutes as a reported figure, not a measured specification.

A Gemini agent in the loop. This is the most interesting part architecturally. Per the blog, Gemini 3.8 Flash runs as a “guiding agent” that “continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded.” If the world changes or the user intervenes, it re-plans. In other words, the pixel generator gets a supervisor that does language-level planning and checks the results. That is how DeepMind claims “reliable causality” and multi-step events, two things a purely autoregressive frame predictor struggles with. Tejas Kulkarni, a research director at Google, told TechCrunch: “What we’ll increasingly see in the future are these hybrid world models.”

The robotics angle

The demo developers in robotics should look at pairs Genie 4 with Gemini 4 Argon and Gemini Robotics. Argon splits a user’s task into steps, and Genie 4 renders each step. DeepMind’s examples include a Universal Robots UR5 arm dispensing liquid with a pipette, and Apptronik’s Apollo humanoid playing checkers move by move. TechCrunch describes this as “a possible way to test robot scenarios in simulation before they happen in the real world.” The Genie model page also says agents such as SIMA can learn new skills inside these worlds.

This is the long-term business case for world models. Physics engines such as MuJoCo or Isaac Sim need every object, material and joint modelled by hand. A learned world model can produce a messy kitchen or a crowded warehouse from a sentence. The catch is that its physics is only plausible, not guaranteed. A policy trained in a world that slowly drifts can learn the drift. DeepMind has published no numbers showing that skills learned in Genie 4 transfer to real robots.

Context: a crowded and expensive field

Genie 4 arrives in the middle of a buying spree. According to TechCrunch, Runway shipped GWM Worlds 2, a playable world model with audio, a couple of months earlier. Fei-Fei Li’s World Labs is being acquired by AMD, Amazon is reported to be in talks to buy Decart, and Rivian has just acquired Third Dimension to bring simulation for self-driving testing in-house. Interactive world generation has moved from research demo to strategic asset. Google’s advantage is integration: the same Gemini models already used for agents and robotics now steer the simulator.

What we don’t know

The announcement is light on numbers. DeepMind gives no resolution, no frame rate, no latency figures and no benchmark scores for Genie 4. One Wikipedia line mentions “up to 30 fps”, but it is unsourced and tagged “citation needed”. The safety section says DeepMind ran internal evaluations before the preview but does not publish them. It also says all generated video and audio will carry a SynthID watermark. DeepMind lists its own limitations clearly: details and object permanence “can still shift or drift over time”, especially in long sessions, after many prompt edits, or while moving through the scene. Complex physical interactions, legible text, spatial accuracy and audio accuracy remain open research problems. Rivas told TechCrunch that “it’s still early days” and that the end-to-end product is not yet settled. There is also a naming oddity: the agent in the loop is Gemini 3.8 Flash, while the robotics demo uses Gemini 4 Argon, and DeepMind does not explain how the two divide the work.

Why it matters for IT

For developers, the important pattern is not the scenery. It is the architecture: a generative simulator under a language-model planner that checks its output frame by frame. The same pattern can produce synthetic test environments for agents, robot-rehearsal pipelines and game or film prototyping without a 3D asset pipeline. Until DeepMind publishes measurements and the API ships, Genie 4 is a well-made preview. If the two-minute coherence and the agent-checked causality hold up for outside testers, simulation for embodied AI will become something you prompt rather than build.