Google DeepMind on October 27, 2026 announced Genie 4, a world model that generates explorable environments in real time, lets users redirect a scene with typed instructions while it runs and, for the first time in the Genie series, produces its own sound, including dialogue spoken by the characters it renders.

Nobody outside a small circle can use it yet. According to Google, Genie 4 has started rolling out as a research preview to "a limited group of trusted testers." The company says it is "coming soon" to Google AI Ultra subscribers in Project Genie and to early access partners in the Gemini API and AI Studio. Neither of the two posts Google published on Tuesday gives a date for that wider release. Neither gives a price, a resolution, a frame rate, a latency figure or a maximum session length either.

The announcement came in two parts. A short post on The Keyword, Google's company blog, carried the headline "We're announcing Genie 4, our new world model for real-time interactive simulation." A longer post on the Google DeepMind blog, titled "Genie 4: A new generation of world simulation" and signed by Raia Hadsell, Diego Rivas and Hang Qi, contains the technical description, the demonstration videos, a section on limitations and a waitlist form for Project Genie. Rivas was also one of three authors of the January post that introduced Project Genie, which listed Rivas as a group product manager at Google DeepMind.

ELI5

Google built a computer program that draws a pretend world on screen, lets people move around inside it, and changes what happens when someone types an instruction such as "make it rain." The new part is that the people in these worlds can now talk and the worlds make their own sounds, which brings the technology closer to the AI video tools Google already offers to advertisers and creators. For now only a small group of testers can try it, everything it makes will carry a hidden AI watermark, and Google has not said when anyone else gets access.

What Google DeepMind says Genie 4 does

A world model in DeepMind's usage is a system that generates an environment frame by frame and keeps it responding to actions, rather than producing a fixed clip. According to the DeepMind post, Genie 4 lets a user "set an event in motion," after which the model "simulates what happens next: how the world responds, how it sounds, and how the people in it move, react, and speak."

The model does not stand alone. Both posts describe it as built on other members of Google's model family. "The Genie 4 model incorporates and builds on capabilities from the Gemini family: Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models provide the complex reasoning needed to help maintain a simulation's coherence and continuity," the Keyword post states.

That dependency matters for anyone following Google's advertising tools. Gemini Omni is the same model family Google has been pushing into creative products since spring. YouTube announced Gemini Omni Flash remixing for Shorts at Google I/O on May 19, 2026, allowing creators to remix up to 10 seconds of an eligible Short. A day later, at Google Marketing Live, Google said Gemini Omni would power video creation inside Asset Studio in Google Ads. On July 16, Omni editing reached Google Vids for paid users only. Genie 4 takes that same visual grounding and puts it inside a system that keeps generating while someone steers.

Google frames the work as a step on a longer research path. "Simulating the world is a crucial step toward AGI, enabling safe training environments for robots and agents, and amplifying creative tools across industries like filmmaking, journalism, gaming, and manufacturing," according to the Keyword post. The DeepMind post puts it slightly differently, saying Genie 4 "advances our ability to create safe, infinitely variable simulations."

Four capabilities, in Google's words

The DeepMind post organises the release around four capabilities. The demonstrations for those capabilities, according to Google, are recordings of real-time interactions in which "users use standard keyboard controls for navigation and camera control, and text prompts to steer the scene as it evolves."

Steering a scene while it runs

The first capability is real-time steering. According to the DeepMind post, users can type instructions to change the weather, move the camera or introduce "a new person or object," and Gemini "will contextually adapt" the instruction to fit the surrounding scene. Gemini also acts between inputs: according to the post, it "monitors and steers the scene to keep the world progressing cohesively."

Something close to this existed on paper before. When DeepMind introduced Genie 3 on August 5, 2025, it described "promptable world events," text instructions that alter a generated world. But when Google opened Project Genie to Google AI Ultra subscribers in the United States aged 18 and over on January 29, 2026, the company said promptable events were not yet part of that prototype. Genie 4 puts text steering at the centre of the product description. Whether it arrives in Project Genie in the same form is not stated.

Physics and people

The second capability concerns realism. According to Google, Genie 4 "simulates real-world physics scenarios, such as aerodynamics, collisions, and more." One demonstration prompt on the DeepMind page describes a macro close-up of a maple seed; another sets a green-striped watermelon in a sunlit courtyard with a weathered stone plinth.

DeepMind reserves its strongest language for human figures. "The most dramatic leap is how Genie 4 simulates people, creating characters that move naturally, respond organically to their environments, and can be interactively steered in real-time," the post states. No measurement accompanies the claim. The post offers no benchmark scores, no evaluation tables and no side-by-side comparison with Genie 3 or with any system from another company.

Sound and speech

Audio is the clearest break with earlier Genie releases. "Genie 4 is our first real-time world model to generate audio natively," according to the DeepMind post. Each world carries ambient sound from its setting, and the post gives examples ranging from "wind in the trees to waves crashing at the beach."

Speech is treated as the harder problem. "The true frontier, however, is human speech," the post states. "Genie 4 introduces interactive dialogue that adapts on the fly." Characters, according to Google, converse and react to user inputs or to events in the world as they happen. A caption under the opening demonstration says that all of the accompanying sounds "are produced directly by Genie 4." Real-time audio accuracy is, however, listed among the model's open research problems further down the same page.

Gemini as an agent inside the loop

The fourth capability describes how Google chains its models together. According to the DeepMind post, "Gemini 3.8 Flash serves as a guiding agent in the loop." While Genie 4 generates a world, Gemini "continuously observes the scene, breaks down goals into a sequence of steps, and reviews incoming frames to track the progress and verify that each step succeeded." When the world shifts or a user intervenes, Gemini changes its plan. Google says this produces "worlds with reliable causality, procedural behaviors, and multi-step reasoning."

Gemini 3.8 Flash has had little public detail of its own. In an interview published on October 21, Sundar Pichai described a run of Flash releases from 3.5 to 3.8 over about three months, whose lessons fed into the post-training of Gemini 4 Argon. The Genie 4 post is one of the few Google documents that assigns 3.8 Flash a specific job.

Gemini 4 Argon and a humanoid playing checkers

DeepMind also connects Genie 4 to its robotics work. "We have also been experimenting with Gemini 4 Argon and Genie 4 to bring our frontier research together with Gemini Robotics," the post states. In the example shown, Argon breaks a user's task into steps, which are passed to Genie 4 as text captions while the simulation renders. The video depicts Apptronik's Apollo humanoid playing a game of checkers. A separate demonstration prompt on the page describes a Universal Robots UR5 arm holding a pipette and dispensing liquid into a lab plate.

The Apollo robot has appeared in Google material before. When Google released Gemini 4 Argon on October 13, it described a "Video-to-Sim-to-Real" pipeline in which Argon watched a video of a person playing cornhole, mapped the environment, calculated joint movements and wrote simulation code. According to Google, a robot built on Apollo then sank seven consecutive shots in an internal trial. That pipeline simulated the world through code; Genie 4 simulates it through pixels. The DeepMind post argues that the two approaches should converge.

Robotics is a long-running line of work at DeepMind. Gemini Robotics, a vision-language-action model built on Gemini 2.0, launched on March 12, 2025, and on June 9, 2026 DeepMind named 15 European startups from nine countries to the first cohort of its Accelerator: Robotics programme. Genie 4's role, in Google's description, is to let AI systems be tested in simulated environments "before bringing them to the real world."

What trusted testers are doing with it

"While Genie 4 is a major research milestone, it remains an early, experimental model," according to the DeepMind post. Google says it has been working with "cohorts of trusted testers across creative industries, journalism, education, gaming, and engineering." The list differs slightly from the Keyword post, which names filmmaking, journalism, gaming and manufacturing. No tester is named in either post, and no company appears as a partner apart from the robot makers whose hardware is depicted.

The DeepMind post groups tester feedback into four uses.

  • Creative direction. Creatives and technologists are testing Genie 4 as a rapid "draft mode" for pre-visualization and ideation. According to Google, filmmakers can step into a live simulation of a scene and run multiple "takes," adjusting camera angles, lighting, dialogue and action as they go.
  • Historical scenarios and immersive photographs. Photojournalists are exploring how "a single 2D archival photograph can become a grounded simulation of that environment in seconds." Teams are also testing how simulated worlds could show short-term environmental events such as a flash flood or a wildfire.
  • Interactive concept design. Game developers and creative technologists are using Genie 4 to move from an idea to an explorable prototype, testing visual styles and interactive mechanics through prompts "before writing a single line of engine code."
  • Simulation. Enterprise partners "across industries" are evaluating Genie 4 for what Google calls rapid "zero-to-one" scenario generation, and identifying gaps the model still has.

The production use case has precedent in Google's own events. At Google I/O in May, director Doug Liman described how AI-rendered environments cut the projected post-production time on the film "Bitcoin" from 40 years to six months, with roughly 150 cities depicted. Liman's team relied on an animatic built from AI-generated stills; what Google describes for Genie 4 is closer to a rough cut that can be walked through and re-staged live.

The gaming use case lands three weeks after Google opened Playground, an experimental browser-based platform for creating games from text descriptions, on October 7. Playground is limited to adults aged 18 and over in the United States, tied to Google AI subscription tiers and testing an integration with Unity. The Genie 4 posts do not mention Playground, and Google has not said whether the two products will connect.

The journalism use case raises questions the posts do not take up. A simulation generated from an archival photograph is, by construction, an invented scene presented in a photographic style. Google's answer, as far as the posts go, is the watermark.

Toward hybrid world models

The final research section of the DeepMind post sets out where Google wants the work to go. "A promising path for future progress lies in hybrid world models, which combine the logical, structured reasoning of code and language models with the rich, sensory representation of pixel-level simulation," the post states.

In that arrangement, Gemini "works behind the scenes to map the structure and plan the narrative timeline, while Genie generates the immediate physical reality - rendering how the world looks, sounds, and responds, pixel by pixel." Google calls the pairing of Genie and Gemini "an early step in this direction." The Argon cornhole pipeline, which wrote simulation code, and the Genie 4 checkers demonstration, which renders pixels from Argon's text steps, illustrate the two sides of that design.

What the announcement leaves out

For a release that leans on the word "unprecedented," the documentation is thin on numbers. Genie 3's launch post on August 5, 2025 gave concrete figures: 720p resolution at 24 frames per second, consistency for "a few minutes," and visual memory reaching back to "one minute ago." When Project Genie launched in January 2026, Google capped each generation at 60 seconds and excluded Google AI Ultra for Business accounts. The Genie 4 posts give none of these parameters. They do not say whether the 60-second cap will change, whether the model runs at a higher resolution than its predecessor or how long a session can last before it drifts.

Several other details are missing:

  • Training data. Neither post describes what Genie 4 was trained on, a question that matters to publishers, photographers and studios whose material could resemble its output.
  • Price and plan. Access is promised to Google AI Ultra subscribers, a tier priced at $99.99 or $199.99 a month, and to early access partners of the Gemini API. No per-second or per-session API price is given.
  • Geography and age limits. Project Genie launched for United States users aged 18 and over. The Genie 4 posts do not say which countries will get access.
  • Compute. Google has not said what serving a real-time, audio-generating world costs. Days before the announcement, Sundar Pichai said Google remains capacity constrained after Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, roughly three times Gemini 3.0's launch figure. How that constraint affects Genie 4's timetable is not addressed.

Google is open about the model's weaknesses. According to the DeepMind post, Genie 4 can maintain coherent environments over short sessions, but "visual details and object permanence can still shift or drift over time - especially as a simulation runs longer, after multiple complex prompt edits, or when navigating throughout the scene." It adds: "Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are active areas of ongoing research." Legible text was also listed as a limitation of Genie 3 in 2025, alongside complex interactions between multiple agents.

Safety, watermarks and labelling rules

Google describes three safety steps. According to the DeepMind post, Genie 4 "includes strict safety guardrails that are consistent with our safety policies across other models," was put through "internal rigorous safety evaluations to prevent the generation of harmful or malicious content" before the preview, and was deployed first to trusted testers to gather feedback before a broader release. The post does not name the evaluations or publish their results. It does not say whether Genie 4 was assessed under Google's Frontier Safety Framework, whose version 3.1 was published on April 17, 2026 and which governs Google's frontier model releases.

The firmest commitment concerns provenance. Every video and audio output from Genie 4 will carry SynthID, Google DeepMind's watermarking system. "To maintain transparency, all video and audio generated by Genie 4 will include SynthID, our imperceptible digital watermarking technology that helps identify AI-generated content across visual and audio modalities," the post states. Google said that as of May 19, 2026 it had marked more than 100 billion images and videos and 60,000 years of audio, and on October 6 it opened the SynthID Detector portal to anyone with a sign-in. Since December 18, 2025, the Gemini app has been able to check uploaded videos of up to 100 MB and 90 seconds for SynthID.

One difference stands out. Google's model page for Gemini Omni says content created or edited with Omni carries both a SynthID watermark and C2PA Content Credentials, the signed metadata standard that records how a file was made. The Genie 4 posts mention SynthID only. Whether Genie 4 output will also carry C2PA metadata is not stated.

The labelling landscape around such output has tightened during 2026. EU AI Act Article 50 transparency obligations became applicable on August 2, 2026, with penalties of up to EUR 15 million or 3% of worldwide annual turnover, and a December 2, 2026 deadline for machine-readable marking by generative systems already on the market. New York's synthetic performer law, in force since June 9, 2026, requires advertisers to disclose fabricated humans in commercial ads, with fines of $1,000 for a first violation and $5,000 for each one after. On July 9, Google added a "How this ad was made" panel to My Ad Center, which discloses AI use automatically for ads built with Google's own generative tools and relies on advertisers to flag AI from outside tools. A model whose headline feature is realistic people who speak produces exactly the kind of content these rules address.

Why it matters for the marketing community

Neither post mentions advertising. Genie 4 is presented as research, with filmmaking, journalism, gaming, education, engineering and robotics as the named fields. The commercial path, though, runs through products advertisers already touch. The visual foundation is Gemini Omni, which Google has placed in Asset Studio, YouTube Shorts and Google Vids within the past six months. The access route is Google AI Ultra and the Gemini API, both paid. The steering layer is Gemini 3.8 Flash, and the robotics demonstrations run on Gemini 4 Argon, which Google has also kept behind paid tiers.

For agencies and production teams, the "draft mode" use case is the most direct. Pre-visualization of a shoot, with camera angles and dialogue adjusted live, is the kind of work Google's testers describe. Whether a simulated take carries enough fidelity to replace a storyboard, an animatic or a test shoot cannot be judged from two posts and a set of curated videos. Google itself says details drift as sessions run longer.

For game studios, Genie 4 joins Playground as a second Google route from text prompt to playable prototype within a month. For publishers, the open questions are training data and the photojournalism use case, where a generated scene based on a real archival photograph would rely on an invisible watermark to distinguish it from documentation.

And for anyone tracking Google's model roadmap, the release confirms a pattern. New capabilities arrive first as a research preview, then reach the most expensive consumer tier and selected API partners, with specifications and pricing filled in later. Genie 4 has completed the first step. The second has no date.

Timeline

Summary

Who: Google DeepMind, in posts signed by Raia Hadsell, Diego Rivas and Hang Qi and on Google's company blog. The model builds on Gemini Omni, Gemini 3.8 Flash and, in robotics demonstrations, Gemini 4 Argon. Trusted testers in creative industries, journalism, education, gaming and engineering have early access; Google AI Ultra subscribers and Gemini API early access partners are next in line.

What: Genie 4, a world model that generates interactive environments in real time, accepts typed instructions to change a scene while it runs, simulates physics and human characters, and generates its own audio including interactive dialogue. All video and audio it generates will carry a SynthID watermark. Google published no resolution, frame rate, session length, benchmark, price or release date.

When: Announced October 27, 2026. The research preview for trusted testers has started; access through Project Genie, the Gemini API and AI Studio is described only as "coming soon."

Where: On the Google DeepMind blog and The Keyword. Future access is through Project Genie, which launched for United States users aged 18 and over in January 2026, and through Google's developer platforms. No country list was given for Genie 4.

Why: Google says simulating the world is "a crucial step toward AGI," providing training environments for robots and agents and tools for filmmaking, journalism, gaming and manufacturing. For the marketing community, the model extends the Gemini Omni line already inside Asset Studio, Shorts and Vids into live, steerable scenes with speaking characters, arriving in a year when EU and New York rules have made disclosure of AI-generated people and media a legal obligation.