Google DeepMind previews Genie 4, an AI world model that adds sound

Google DeepMind announced Genie 4 on 27 October 2026. Genie 4 is its new world model, an AI system that generates an explorable environment in real time as a user moves through it. The headline change is sound: Genie 4 generates ambient noise, music and interactive dialogue together with the picture. Google calls it “our first real-time world model to generate audio natively”. A research preview is open to a limited group of trusted testers. Wider access is planned for Google AI Ultra subscribers and early developer partners.

The other changes are live scene editing by text, a longer stretch of consistency and better physics. Each of those claims comes from Google. So far the public evidence is a set of selected demonstrations, not measured results.

What Genie 4 adds over Genie 3

Google released Genie 3 as a research preview a little over a year ago. According to Google and early coverage, Genie 4 improves on it in four areas.

• Native audio. Every world generates ambient sounds, music and interactive dialogue on the fly. DeepMind says the model is “natively multimodal”, meaning audio and video are generated together rather than added afterwards. According to Ars Technica, users can specify audio elements in the prompt, and dialogue is meant to adapt as the simulation changes.
• Live steering. Users can type instructions mid-simulation to move the camera, change the weather or introduce a new person or object. TechCrunch notes that Genie 4 allows free-form edits of this kind in real time.
• Longer consistency. Genie 4 can keep a scene consistent for up to two minutes, double its predecessor’s limit. The Verge puts it as experiences of about two minutes, up from one minute for Genie 3.
• Physics and people. DeepMind says Genie 4 handles aerodynamics and collisions better. It also says generated people move more naturally and respond to a changing environment. Google’s model page describes environments where “objects collide, water flows, and trees sway in the wind”.

Google’s model page calls Genie 4 “our most realistic and interactive world model yet”. TechCrunch describes its output as the most photorealistic and interactive worlds yet. That assessment draws on Google’s demos and a press briefing, not on hands-on testing.

How Gemini directs the simulation

Genie 4 is built on capabilities from Google’s Gemini models. In Google’s description, Gemini Omni “anchors simulations in real-world appearance and behavior”. Other Gemini models supply the reasoning that keeps a simulation coherent and continuous.

Gemini 3.8 Flash is the model that plays the director. It watches the generated frames, breaks a goal into steps and checks each new image to see whether a step succeeded. It changes its plan when the user steps in or the scene shifts. When a user types an edit, DeepMind says, Gemini adapts the prompt to fit the surrounding scene. Between user inputs, Gemini also monitors and steers the scene to keep it progressing cohesively. Ars Technica compares this to the way reasoning models generate extra internal prompts to reach the right output.

Tejas Kulkarni, a research director at Google, described where this design is heading during a briefing:

“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.”

Google has not published a full architecture. Separate benchmark scores for Gemini models do not measure the complete simulation.

Specs: what Google has and has not disclosed

The launch materials give capabilities, not technical figures. Kingy AI reviewed the launch and product pages. It found no numerical Genie 4 benchmark results and no stated frame rate, resolution, parameter count or API price.

Item	What the sources say
Model type	Real-time interactive world model
Controls	Keyboard controls for navigation and camera; text prompts to steer
Audio	Ambient sound, music, interactive dialogue, generated natively
Scene consistency	Up to two minutes, double Genie 3
Planning model	Gemini 3.8 Flash
Resolution, frame rate, latency	Not disclosed
Parameters, training data	Not disclosed
Weights	No public release found
Watermarking	Google’s invisible SynthID on generated video and audio

The Kingy AI review also found other gaps for developers. Google has not documented a game-engine export, a public save-state interface or microphone input for Genie 4. Google’s own demos come with a disclaimer that the results are illustrative and the sequences shortened. The disclaimer does not give attempt counts or failure rates.

Access: testers now, Ultra and API partners next

Genie 4 is available now as a research preview for trusted testers. DeepMind says the preview is for gathering feedback before release to consumers and developers. It names two next steps:

• Google AI Ultra subscribers, through Project Genie, Google’s tool for generating interactive worlds from text or image prompts.
• Early access partners of the Gemini API and Google AI Studio.

Google calls these routes “coming soon”, and TechCrunch reports them for the coming weeks. Ars Technica reports that there is no specific timeline yet and that API pricing is not yet available. DeepMind has opened a waitlist for Genie 4 in Project Genie.

Ars Technica notes that Genie 3 never expanded beyond the limited Project Genie web experience. It adds that Google may eventually extend Genie 4 to its AI Pro and AI Plus plans. Kingy AI cautions that a Project Genie subscription is not an API plan. It found no public Genie 4 endpoint, model ID, SDK example or price.

What it means for builders

Google lists several uses for the model:

• Training robots and agents in simulated environments.
• Game prototyping and interactive learning.
• Pre-visualisation for film.

The company names filmmaking, journalism, gaming and manufacturing as industries it hopes to serve.

Film and creative prototyping is the use that trusted testers are pointing to, according to TechCrunch. Diego Rivas, a Google DeepMind engineer, said at the briefing:

“The real time aspect becomes most valuable in the areas where keeping a state of flow in the creativity space becomes incredibly useful. For example, in the ideation phase, in the storyboarding phase — anytime you’re kind of in need of rapid ideation.”

Rivas added that Gemini Omni could eventually turn these prototypes into creative assets. He said those capabilities are still a ways off.

Robotics is the second use. In a research experiment tied to Gemini Robotics, a Gemini 4 Argon model broke tasks into steps, and Genie 4 simulated those steps in real time. The demo shows an Apptronik Apollo humanoid robot playing checkers. Kingy AI notes that this demo does not establish how a physical robot would perform or whether skills transfer from simulation to the real world.

For teams building on APIs, there is nothing to integrate yet. Kingy AI notes that production planning needs a documented endpoint, quotas, pricing and terms. None has been published so far. Kingy AI also cautions that a generated world is not a game engine. A convincing scene does not prove there is a reusable 3D model, a persistent object store or a deterministic physics simulation underneath it.

Limits, risks and open questions

Google acknowledges clear gaps. Details and objects can drift the longer a simulation runs. The team also has not worked out how to make audio accurate or in-world text legible. Objects can disappear after several edits. DeepMind still treats complex physical interactions and long-term consistency as research problems. Rivas stressed that it is “still early days” for Genie 4 and that the end-to-end product experience is not yet clear.

Outside critics raise doubts in two areas, games and robotics.

• Games. When Project Genie launched earlier this year, shares of Unity, Roblox and Take-Two fell sharply. Industry figures called the sell-off overblown. Andrew Green, chief executive of the studio Storygrounds, said of complex games: “You can’t prompt your way to that. Not yet. Not for a long, long time,” as reported by Pocketgamer.biz.
• Robotics. The AI writer Ben Dickson argues that a generated world has no ground truth to check against. Errors in the model could therefore compound when robots are trained inside it.

Testing the audio claims will take more than one clip. Kingy AI suggests interactive tests across many sessions. Does a door’s sound land at the right moment? Does a voice move with a character who walks away? No independent tester has published results yet. The Verge’s reporter had not been able to try the model, and Kingy AI has not tested it hands-on.

The competition

Genie 4 arrives a couple of months after Runway released GWM Worlds 2, a playable world model that also generates audio. The field is consolidating:

• AMD is acquiring Fei-Fei Li’s World Labs, for $8.2 billion according to Trending Topics.
• Amazon is in talks to buy Decart.
• Rivian has acquired Third Dimension to build simulation for self-driving testing in-house.
• Nvidia is pursuing world models for robots and self-driving cars through its Cosmos platform.

What to watch next

• Developer access. Watch for Gemini API and AI Studio documentation: a model ID, quotas, latency figures and pricing.
• The Ultra rollout. Watch when Genie 4 reaches Project Genie for Ultra subscribers, and how tightly Project Genie limits sessions.
• Independent tests. Watch for hands-on reports on audio sync, drift over two-minute sessions and how the model responds to edits. None exists yet.
• Robotics evidence. Watch for any results that show whether training in Genie 4 transfers to physical robots.