Google’s Genie 4 can now build a world you steer in real time, with sound

Genie 4 is open, for now, to a short list of trusted testers and to nobody else. Google DeepMind released the world model on October 27 as a research preview, and says it will reach Google AI Ultra subscribers through Project Genie and early-access partners of the Gemini API and Google AI Studio “soon.” A waitlist is linked from the company’s page, and no date has been given for any of the wider steps.

Stay ahead on AI: Join 600,000+ professionals who read The Deep View, the daily newsletter that keeps you up to date with artificial intelligence.

Subscribe to The Deep View →

We may earn a commission if you sign up.

What the testers get is a generated, explorable scene that responds as it runs: movement from standard keyboard controls, changes to the scene from typed prompts, and a soundtrack produced alongside the picture. Everything about quality and behavior comes from the company’s own posts and a single trade-press briefing, since no outside group has yet published tests, so the claims below are DeepMind’s and the gaps in what it has published are noted where they matter.

Keyboard for movement, text for everything else

In its launch post, authored by Raia Hadsell, Diego Rivas and Hang Qi, DeepMind describes Genie 4 as a real-time interactive world simulation. Users navigate with ordinary keyboard controls and steer the scene with text prompts, and the post says Gemini adapts requests such as changing the weather, moving the camera or adding people and objects so they fit the world already on screen.

The companion Google blog post adds the intended uses: training environments for robots and agents, and creative work in filmmaking, journalism, gaming and manufacturing. The same post frames simulation as a step toward artificial general intelligence, an aspiration rather than a measured result.

On duration, the sources differ in wording. TechCrunch reported that a scene stays consistent for up to two minutes, double the limit of Genie 3. Google’s earlier Project Genie announcement put Genie 3 generations at 60 seconds, which fits that doubling. That earlier release reached the public on January 29, 2026, as a web app limited to Google AI Ultra subscribers in the United States aged 18 or older, and it ran on Genie 3 together with Nano Banana Pro and Gemini, and Google names the same subscriber tier, through Project Genie, as the next audience for Genie 4. DeepMind’s own Genie page is looser, saying the model supports “a few minutes of continuous interaction” and is not designed for extended sessions.

Sound generated with the picture

Audio is the main addition. DeepMind calls Genie 4 natively multimodal and says it produces music, ambient sound and interactive dialogue as the world runs, and its post calls it the first real-time world model to generate audio natively. The claim is the company’s, and the preview’s closed access means outside testers have not yet checked it.

DeepMind is careful about the limit. On the model page it states that real-time audio accuracy “remains an area of ongoing research and improvement.” TechCrunch reported that the team has not yet worked out how to guarantee audio accuracy or legible text inside the generated world.

Gemini as director, and a robot test

The architecture leans on the Gemini family. According to DeepMind, Gemini Omni anchors scenes in real-world appearance and behavior, other Gemini models handle reasoning to keep a simulation coherent, and Gemini 3.8 Flash acts as a guiding agent that monitors frames, breaks goals into steps and verifies progress. Tejas Kulkarni, a research director at Google, told TechCrunch that hybrid world models will increasingly define what comes next, a phrase DeepMind’s post echoes in proposing models that combine language reasoning with pixel-level simulation. Kulkarni’s point, as TechCrunch relayed it, is that the future of these systems lies in pairing a language model’s planning with a generator’s rendering, not in a single network doing both.

One experiment moves beyond games. Per DeepMind and TechCrunch, a model called Gemini 4 Argon, working in a Gemini Robotics setting, split tasks into steps that Genie 4 simulated in real time, which suggests robot scenarios could be rehearsed in simulation before any hardware moves. TechCrunch reported that trusted testers also point to film ideation and storyboarding.

Rivas, a DeepMind engineer and one of the post’s authors, told TechCrunch that the real-time aspect matters most in areas where a user needs to keep a state of flow, such as game-like exploration, and that it is still early and the end-to-end product experience is not yet clear. That candor is notable in a launch that otherwise leans on large ambitions, and it matches the limits listed in the company’s own post.

What DeepMind lists as unsolved, and what it has not published

The launch post says visual details and object permanence can drift over longer sessions or after multiple edits, and that physical interactions, spatial accuracy, legible text and long-term consistency remain research problems. It adds that Genie 4 went through internal safety evaluations and that all generated video and audio will carry SynthID watermarks.

Two figures a reader might expect are absent. None of the DeepMind or Google pages read for this report states the output resolution or the frame rate, and no Genie 4 model card turned up; the model cards that surfaced in a search belong to other Gemini models. Until DeepMind publishes them, claims about how sharp or smooth the worlds look rest on the company’s demonstrations. The next measurable test is the Ultra rollout: how long a scene holds together in outside hands, and whether the audio holds up, will be visible once subscribers can use it.

This article was produced with the assistance of AI and reviewed by Morning Overview editors prior to publication.

Stay ahead on AI: Subscribe to The Deep View →

We may earn a commission if you sign up.