DeepMind’s Genie 4 Adds Sound to Worlds That Last a Few Minutes

Google DeepMind’s newest world model can now be heard as well as seen. Genie 4, announced on 27 October, generates an explorable scene in real time together with its own sound: ambient noise, music and characters who speak and react to the user. DeepMind’s own model page also sets the limit, saying it “can support a few minutes of continuous interaction, rather than extended sessions”.

That pairing sums up the release. The London-based lab is claiming a jump in realism, with better physics, more natural people and native audio, while keeping the model behind a waitlist and publishing no test results. DeepMind’s announcement says Genie 4 is rolling out “as a research preview for trusted testers”, and that it is “coming soon” to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and Google AI Studio. Neither page gives a date or a price.

Sound is the new part, and speech the hardest

A world model, in this sense, is an AI system that generates the next moments of a scene in response to what a user does, instead of playing back a fixed video. In Genie 4, users move and turn the camera with standard keyboard controls and type prompts to change the scene as it runs, for example “changing the weather, moving the camera, or introducing a new person or object”. DeepMind says Gemini adapts each prompt to fit the surrounding scene and keeps steering the world between a user’s inputs.

“Genie 4 is our first real-time world model to generate audio natively,” DeepMind writes. Every world carries the sound of its setting, from wind in trees to waves on a beach. The lab calls human speech “the true frontier”: characters hold “interactive dialogue that adapts on the fly”, reacting to the user or to events in the world.

DeepMind also says Genie 4 simulates physics such as “aerodynamics, collisions, and more”, and that “the most dramatic leap” is in how it renders people, who move naturally and can be steered in real time. Google’s Keyword blog post repeats the claim of “unprecedented realism”. The evidence offered is a set of demo clips.

Gemini does the planning in the background

Genie 4 leans on Google’s Gemini models. DeepMind says Gemini Omni “anchors the simulations in real-world appearance and behavior”, while Gemini 3.8 Flash acts as “a guiding agent in the loop”: it watches the scene, breaks a goal into steps and checks incoming frames to confirm each step worked, changing its plan if the user steps in.

The lab has also paired Genie 4 with Gemini 4 Argon as part of its Gemini Robotics research. In one demo, Argon breaks a task into steps that are fed to Genie 4 as text captions while it renders a simulation of Apptronik’s Apollo humanoid playing checkers. On the same day, Google put a different robotics model, Gemini Robotics ER 2, in front of business customers on Google Cloud. Genie 4’s robotics work is still a research experiment.

TechCrunch reported from a Google briefing that research director Tejas Kulkarni sees the future in “hybrid world models, where you have a combination of Gemini writing code and text, and Genie ingesting that in real time”. DeepMind’s blog uses the same term for its long-term plan.

Testers include journalists, and every output is watermarked

DeepMind says it has worked with “cohorts of trusted testers across creative industries, journalism, education, gaming, and engineering”. Filmmakers are using it as a “draft mode” for pre-visualisation, game developers to prototype ideas, and enterprise partners for scenario generation. Photojournalists, it says, are exploring how “a single 2D archival photograph can become a grounded simulation of that environment in seconds”.

DeepMind says it ran “internal rigorous safety evaluations” before the preview and released the model to testers first to gather feedback. All video and audio that Genie 4 generates “will include SynthID”, Google’s invisible watermark for identifying AI-made content.

DeepMind lists its own limits

DeepMind calls Genie 4 “an early, experimental model”. Its blog says that while the model keeps environments coherent over short sessions, “visual details and object permanence can still shift or drift over time”, especially as a simulation runs longer, after several complex prompt edits, or as a user moves around the scene. The model page says it supports “a few minutes of continuous interaction”; TechCrunch reported that Google put the figure at up to two minutes, double the limit of Genie 3.

The model page adds that Genie 4 “is currently unable to simulate real-world locations with perfect accuracy”, and that legible text “is often only generated when it’s in the input world description”. Google has published no benchmark scores, resolution, frame rate or price.

A few minutes of world, and nothing to measure it against

The short session limit is the most useful number in the release, and DeepMind deserves credit for printing it. A couple of minutes is long enough for a film storyboard or a quick game idea. It is short for training a robot or an AI agent on a long, multi-step task, which is the use DeepMind leans on when it talks about the road to AGI. Drift that grows with every prompt edit cuts against the steering that is meant to be Genie 4’s selling point.

Without benchmarks, resolution or frame rate, the realism claims rest on clips DeepMind chose itself.

That is a weaker evidence base than recent rivals offered. Tencent shipped Hy-Nomas with open weights and a 280-case test set, though no scores. Noiz AI and HKUST published a paper and figures for Helix World, which also generates sound with the picture. Odyssey released numbers for Odyssey-3, all its own. Genie 4 may be ahead of all three, but no outsider can check yet.

The journalism use deserves the closest watch. A photo turned into a moving, talking scene is a reconstruction, not a record. SynthID helps only if platforms and newsrooms check for it, and a watermark says a clip was made by AI, not whether the scene it shows is faithful. For any newsroom or compliance team, provenance rules will need to treat such output as synthetic media from the first frame.

The waitlist is the next marker

Google’s pages say only that Genie 4 is “coming soon” to AI Ultra subscribers and developer partners; TechCrunch reported a rollout “in the coming weeks”. The open questions are the price, what session length paying users will get, and whether DeepMind publishes an evaluation before the doors open wider.

Sources

1. Google DeepMind: “Genie 4: A new generation of world simulation”, Raia Hadsell, Diego Rivas and Hang Qi, 27 October 2026.
2. Google: “We’re announcing Genie 4, our new world model for real-time interactive simulation”, 27 October 2026.
3. Google DeepMind: Genie 4 model page, accessed 27 October 2026.
4. TechCrunch: “Google’s Genie 4 world model adds audio, more realistic worlds, and real-time steering”, Rebecca Bellan, 27 October 2026.

Related stories

• Tencent Hunyuan Open-Sources Hy-Nomas, an Interactive World Model
• Noiz AI and HKUST Release Helix World, a Real-Time World Model Whose Sound Turns With the Camera
• Odyssey Opens Its Odyssey-3 World Model to the Public; Its Headline Physics Score Is a Best-of-Eight, and Every Benchmark Is Odyssey’s Own