Genie 4 Adds Sound and Real-Time Steering to DeepMind's World Model

Google DeepMind announced Genie 4 on October 27, 2026, and the addition likely to matter most is sound. The world model, which generates an interactive scene from a prompt or a still image, now produces audio with the picture: ambient sound that matches the setting, music, and dialogue that answers what the user does. DeepMind calls the release a research preview for trusted testers, not a product anyone can sign up for today.

Related reading: Google's Gemini 4 Argon Reaches Trusted Cyber Defenders First, With a 1 Million-Token Output Limit · Google's Gemini 3.8 Live Reasons Out Loud While It Talks.

Access will widen slowly. The company says Genie 4 will reach Google AI Ultra subscribers inside Project Genie and early-access partners using the Gemini API and Google AI Studio in the coming weeks, and a waitlist is open, The Verge reported. The second change is control. A user can type an instruction while a world is running to change the weather, move the camera, or introduce a person or object, and the model rewrites the surroundings to fit instead of snapping back to a starting frame.

What sits under the demo

Genie 4 builds on the Gemini family rather than replacing it. DeepMind's announcement says Gemini Omni anchors each simulation in real-world appearance and behaviour, while other Gemini models supply the reasoning that keeps a scene coherent from one moment to the next. The system is natively multimodal in the company's phrasing: audio is generated as part of the world, not layered on afterwards.

World models are usually separated from ordinary video generators by interactivity. The output responds to input instead of playing a fixed sequence, and the steering claim is what makes that difference testable. If a user can move the camera and the scene reorganises around the new viewpoint, the model has something like an internal model of the space. If it only produces a plausible next frame, the word is doing less work.

The less visible change is an agent inside the loop. Gemini 3.8 Flash observes the generated scene, breaks a goal into steps, reviews frames and checks whether the world is still doing what was asked. DeepMind credits that loop with extending how long a scene holds together, to roughly two minutes. TechCrunch reports that is double the limit of Genie 3. Users do not get a two-minute video file out of it. They get a world that can stay stable for about that long while someone steers it.

How much work that loop is doing is hard to judge from outside. An agent that checks frames can catch a scene drifting away from its prompt, but the same agent is part of the generation pipeline. DeepMind has not published numbers on how often it intervenes or how many attempts a finished clip takes. The architecture is the most interesting part of the announcement and the least documented.

Four claims, one demo reel

DeepMind groups the improvements into four areas.

• Steering. Natural-language instructions change weather, camera position and the contents of a scene mid-run, with the model adapting each prompt to what is already on screen.
• Physics and people. The company claims better aerodynamics and collisions, and figures that move like humans and react to their surroundings.
• Audio. Ambient sound, music and interactive dialogue generated alongside the visuals and adjusted on the fly.
• Multi-step events. Gemini reasoning plans a sequence and checks it, instead of generating each moment in isolation. In one robotics demonstration, a Gemini 4 Argon model broke a task into steps for Genie 4 to simulate, with an Apptronik Apollo robot playing checkers.

Those are company claims. The supporting evidence in the launch material is a set of short clips, and Ars Technica built its coverage around them. The clip that opens the footage shows a man in a brown suit playing a saxophone on a platform above a misty lake. The water moves, the instrument has weight, and the sound belongs to the scene rather than sitting on top of it.

Audio is the harder claim to assess from a press release. Dialogue that stays in sync with a moving mouth is a different problem from ambient sound that matches a forest, and DeepMind groups both under one heading. The clips suggest the company can do both in short bursts. They do not show a conversation that lasts.

Who is meant to use it

DeepMind lists film and television pre-visualisation, storyboarding, game prototyping, photojournalism and robotics simulation. Its engineers describe the film case in the most detail. Diego Rivas, an engineer on the project, told TechCrunch that speed matters when it keeps a creative team in flow: "The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful ... in the ideation phase, in the storyboarding phase."

The photojournalism example is stranger. DeepMind says testers have taken a two-dimensional archival photograph and turned it into a scene someone can walk through. That is a new kind of historical reconstruction and a new set of questions about what a photograph is evidence for. The launch material does not say how the model handles the parts of the scene the photograph never contained, which is most of them.

For game developers, the pitch is testing a visual style, a material or a mechanic before anyone writes engine code. Studios already prototype scenes before committing to production; a generated world could shorten that step. Whether a studio would plan around a world it cannot inspect at the code level is a separate question, and one the preview does not answer.

DeepMind also names enterprise scenario generation, the phrase companies use for building a visual mock-up of a process before committing money to it. The announcement does not show what those scenarios look like, which leaves the claim hard to assess. Whatever the setting, the value depends on fidelity, and fidelity is exactly what the limitations list puts in doubt.

Tejas Kulkarni, a research director at DeepMind, described the longer arc in an interview with TechCrunch. His version has Gemini models writing code and text while Genie ingests the output and simulates the result, with the world model acting as a renderer and a physics stand-in rather than the whole system.

The money moving into world models

Genie 4 arrives into a market that has spent the past year buying world-model companies. Runway released its second world model, GWM Worlds 2, a couple of months before this launch, according to TechCrunch. AMD agreed on September 28, 2026 to buy Fei-Fei Li's World Labs for $8.2 billion, Amazon is in talks to buy Decart for around $7 billion according to the Wall Street Journal, and Rivian acquired the world-model startup Third Dimension on October 22 to speed up self-driving simulation.

The pattern is consolidation around simulation as a training input. A world model can produce labelled experience for a robot or a car without miles of real driving, which is why the buyers include a chipmaker, an automaker and, reportedly, a cloud provider.

Bloomberg framed the Genie 4 launch as a model that "rattled video game stocks" in its headline on October 27. The story is paywalled and the headline carries no number, so the size of any market move is not established by the material available here. What the headline does show is that investors now read a world-model release as news about game publishers as much as about research labs. They may be getting ahead of the technology; the reaction still says something about expectations.

What we'll increasingly see in the future are these hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction. — Tejas Kulkarni, Google DeepMind research director, to TechCrunch

What it still gets wrong

DeepMind's own limitations list is specific. Visual details and object permanence can drift over longer sessions, after several complex prompt edits, or while the camera moves. Complex physical interactions, audio accuracy, legible text inside a world, spatial accuracy and long-term consistency remain open research problems. TechCrunch reports that the team has not solved audio accuracy or legible text.

The safety disclosure is brief but concrete. DeepMind says all video and audio Genie 4 produces carries SynthID, its imperceptible watermark, and that it ran internal safety evaluations before opening the preview. Watermarking helps platforms identify generated material if they check for it. It does not stop anyone from recording a screen.

What to watch

The first test is delivery. A launch window is not a date, and the gap between a preview for testers and a product for subscribers can run long. Watch whether AI Ultra subscribers get Genie 4 on the schedule DeepMind described, and whether API access arrives with the same capabilities or a reduced set.

The second is independent evaluation. Scene consistency over two minutes is the kind of claim that outside testers can measure, and the interesting result will be how quickly a world falls apart when someone stops steering it the way the demo does. Audio sync and legible text are easier to check and harder to fake. The effect of long prompts is harder still: DeepMind says drift gets worse after several complex edits, which is precisely how a film team would work. A system that works for one clean take and degrades on the fifth revision has a narrower market than the launch suggests.

The third is what buyers do. If game studios start using generated worlds in pre-production, the Bloomberg headline will look prescient. If they treat it as a storyboard tool and nothing more, the market reaction was noise. Robotics is the quieter case: a simulator good enough to train a policy would matter more than any demo clip, and nothing in the launch material shows that yet.

For now, DeepMind has shown a better demo than its last one, with sound. It has not shown that the system holds up when someone outside the company is driving. The API preview and the first wave of testers will be the first real evidence either way.

Cover image: a Genie 4 demo world generated by Google DeepMind, from launch footage published by Ars Technica.