Google DeepMind's Genie 4 Builds Playable Worlds That Talk Back

Type a prompt, and a city street appears. Press the arrow keys, and you walk down it. Type "make it rain," and the sky darkens while you are still moving, with the sound of rain arriving at the same moment. That is the pitch behind Genie 4, the world model Google DeepMind announced on Tuesday, October 27, as a research preview for a small group of trusted testers.

Genie 4 builds interactive scenes in real time and, for the first time in the series, generates audio alongside the picture. Music, background noise and spoken dialogue all come from the same system, according to Google DeepMind. It is one of the clearest signs in our AI coverage this fall that Google sees world models as a product, not only a lab experiment.

Genie 4 adds sound to worlds that used to be silent

Earlier Genie releases produced moving pictures with no audio. Google DeepMind's announcement calls Genie 4 its "first real-time world model to generate audio natively," which means the soundtrack is not stitched on afterward by a separate tool. If a character speaks, the voice is produced as part of the scene.

Visual quality also moved up. DeepMind describes sharper photorealism, more fluid movement, and better simulation of things like collisions and air resistance. People inside the generated worlds walk and react more naturally than before, and users can steer them as the scene plays out.

TechCrunch reports that Genie 4 can keep a scene consistent for up to two minutes, double the limit of Genie 3, which Google previewed in August 2025. Two minutes sounds short, but in world models consistency is the hard part. A scene that forgets where a doorway was after 30 seconds is not much use to anyone.

Gemini acts as the director while Genie draws the frames

The most practical change is how scenes respond to plain language. While a simulation runs, a user can type an instruction such as changing the weather, moving the camera or adding a person or object. Movement still works through standard keyboard controls, much like a PC game.

Behind those edits, DeepMind says Gemini 3.8 Flash watches the scene, breaks a goal into smaller steps and checks individual frames to confirm each step happened. Google describes Gemini as the planner and Genie as the engine that renders what the plan calls for. The appearance and behavior of objects are anchored by Gemini Omni, according to the announcement, while other Gemini models handle reasoning about continuity.

That split of labor matters because the same pattern shows up across the industry. Language models are increasingly used as supervisors for other systems, a theme we covered in our explainer on how AI agents work.

Ultra subscribers and API partners are next in line

For now, only trusted testers can use Genie 4. Google says it will reach Google AI Ultra subscribers inside Project Genie, the company's world building tool, and early access partners of the Gemini API and Google AI Studio. TechCrunch puts that rollout in the coming weeks. DeepMind's own page says only "coming soon" and lists no date. A waitlist for Project Genie is open.

There is no standalone price. Access runs through the Ultra subscription, which sits at the top of Google's consumer AI plans. That tiering is consistent with how Google has been rationing its newest models, as seen when it moved free Gemini users to Flash Lite.

On safety, Google says Genie 4 went through internal evaluations before the preview and follows the same policies as its other models. Every video and audio output will carry a SynthID watermark, Google's invisible tag for identifying AI generated media.

Filmmakers and robot trainers are the first real customers

Google's examples lean toward professional use. Trusted testers have used Genie 4 for film previsualization, running several "takes" of a scene inside a live simulation before anything is shot. Others turned a single archival photo into a space they could walk through, or sketched interactive prototypes from text prompts.

Diego Rivas, a Google DeepMind engineer, told TechCrunch the real time aspect is most valuable for ideation and storyboarding. He also cautioned that it is still early days for the model.

Robotics is the other target. In a Gemini Robotics experiment described by TechCrunch, a model called Gemini 4 Argon broke a task into steps that Genie 4 then simulated live, a way to rehearse robot behavior before trying it on hardware. Training robots in simulated worlds has become a race of its own, with OpenAI also signaling plans for a humanoid robot demo in 2027.

Competition is crowding in. TechCrunch notes that Runway released its GWM Worlds 2 model with audio a couple of months ago, AMD is acquiring Fei-Fei Li's World Labs, and Amazon is reportedly in talks to buy Decart for around $7 billion. Meta, meanwhile, has been pushing its own frontier work after the Muse Wonder launch.

The weaknesses Google lists on its own page

Google is unusually direct about what does not work yet. Visual details and object permanence can drift, especially in longer sessions, after several edits, or while moving through a scene. Complex physical interactions, accurate real time audio, readable in world text, spatial accuracy and long term consistency are all described as open research problems.

Tejas Kulkarni, a Google research director, pointed TechCrunch toward what comes after this: "hybrid world models" that mix Gemini written code and text with Genie's pixel level simulation. If that works, the next version may not just draw a world you can walk through, but one that follows rules you can actually program.