The worlds that AI builds just learned to talk; whether they can hold a conversation long enough to matter is the catch.
On October 27, 2026, Eastern Time, Google DeepMind released Genie 4, its new world model, as a research preview open first to trusted testers. The launch post, "Genie 4: A new generation of world simulation," is credited to Raia Hadsell, Diego Rivas, and Hang Qi. Sundar Pichai marked the day on X: "Excited to share Genie 4 today, our most realistic and interactive world model: steer dynamic environments in real time, shaping events, camera angles, and actions."
The official positioning, per the blog, is that Genie 4 "unlocks unprecedented realism and improves real-time interactive simulation." But the real headline is not the visuals. It is the sound. Genie 4 is, by DeepMind's own account, the first real-time world model to generate audio natively: music, ambient sound, and interactive character dialogue are produced together with the world itself, not dubbed onto a rendered video afterward. Two lines from the blog belong together: "Every world carries the ambient sound of its own setting, from wind in the trees to waves crashing at the beach," and "The true frontier, however, is human speech." Characters can now respond in real time to what you type and to events in their world. The worlds built by AI have, for the first time in this product line, a voice.
DeepMind also wrote something larger into the announcement: "Understanding and simulating the world is a crucial aspect of building AGI." That frames Genie 4 precisely: not merely a toy you can walk around in, but a component of the path toward AGI in DeepMind's roadmap. A note on scope: this piece is a launch deep dive, not a world-model comparison; Runway's GWM Worlds 2 appears only as a reference point, and a full multi-model scoring comparison is a separate article.
Four Advances: Directing, Physics, People, Sound
Google frames the progress over the previous generation as four advances.
The first is real-time steering. You navigate with a keyboard and rewrite the world in natural language while the simulation runs: change the weather, move the camera, add a new character or object. Gemini adapts your prompt to fit the scene, and between your inputs it keeps monitoring and steering the scene so it stays coherent. The world does not wait for your next command; it keeps performing on its own.
The second is physics. The blog says Genie 4 simulates real-world interactions such as aerodynamics and collisions more accurately, which decides whether a simulated result can be trusted for robot testing or game prototyping.
The third is character motion, which Google itself calls the most dramatic leap: "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 plastic way people moved in earlier world models was always the easiest tell; this is where the biggest jump is claimed.
The fourth is audio. The first three generations of the Genie line were silent; this is the first where sound grows inside the world.
Beyond the four, Google lists a separate dimension of reasoning: Gemini models plan multi-step events, so scenes unfold along chains of cause and effect rather than reacting frame by frame. It is the deepest difference between Genie 4 and a video generation model.
A Gemini Inside the Loop
Genie 4 runs on a hybrid architecture. In the official formulation, "Gemini Omni anchors the 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 most interesting role belongs to Gemini 3.8 Flash, a guiding agent inside the loop: as Genie 4 generates a world in real time, it continuously observes the scene, breaks goals into steps, and reviews incoming frames to verify that each step succeeded. When the world changes or the user intervenes, it adjusts the plan on the spot. Google says this produces worlds with "reliable causality, procedural behavior, and multi-step reasoning."
According to Chinese financial outlet Cailianshe, the team tried putting Gemini 4 Argon into the loop, but the thing actually driving the real-time experience is the lighter Gemini 3.8 Flash; the strongest model is not necessarily the right conductor.
There is a vivid official demo: Gemini 4 Argon, the model Google released publicly earlier this month, decomposes a task into steps, and Genie 4 renders their execution in real time, with an Apptronik Apollo humanoid robot playing checkers while the steps appear as captions. Another demo has a Universal Robots UR5 arm dispensing liquid onto a lab board and triggering tiny electric sparks. TechCrunch reads these as a path toward testing robot behavior in simulation before executing it in the physical world.
Tejas Kulkarni, Google's research director, put the thesis more sweepingly at a press briefing: future hybrid world models will combine "Gemini writing code and text, and Genie ingesting that in real time and simulating reality at different levels of abstraction." The world model is no longer a standalone video engine; it is a simulation system with language models as its hands and brain.
Two Minutes: A Real Number With Mixed Accounts
According to TechCrunch and The Verge, Genie 4 can hold a consistent scene for roughly two minutes, double the previous generation.
That "two minutes" needs three layers of qualification. First, DeepMind's blog never quantitatively endorses the figure; it concedes only that visual details and object permanence still drift over long sessions. "Doubled to about two minutes" is the press account, not an official promise. DeepMind research engineer Hang Qi put it bluntly: even at the two-minute ceiling, the picture is not fully consistent throughout.
Second, "about one minute for the previous generation" is a media simplification. The official Genie 3 blog post says 720p resolution, 24fps real-time navigation, and consistency over "several minutes," with visual memory retrievable for about one minute. The figure that actually matches "60 seconds" is the Project Genie prototype from January 2026, built on Genie 3, whose single-generation limit was about 60 seconds. Against that baseline, Genie 4's roughly two minutes is a clean doubling.
Third, two minutes is not yet something to sneer at. DeepMind product manager Diego Rivas told media plainly: what you can do in two minutes is ultimately limited, and the team is still in research exploration, not final product development. The lineage puts the slope in view: Genie 2 in December 2024 was non-real-time, 360p, with single interactions of 10 to 20 seconds; Genie 3 in August 2025 brought real-time generation and interaction; Genie 4 adds native audio and a Gemini reasoning loop. At every generation, the slope of that curve says more than any single number.
The slope has a concrete face. Genie 3's visual memory meant that if you drew graffiti on a wall, walked away, and came back, the graffiti was still there, and Genie 3 also supported Promptable World Events, injecting occurrences into the world by text. What Genie 4 adds on top of that foundation is a director: scenes now advance along planned, multi-step chains of cause and effect, watched over by Gemini in the loop.
Rivals on Stage, Giants Buying the Theater
Two months earlier, on September 3, 2026, Runway released GWM Worlds 2, the most direct competitor, also as a research preview: continuous 720p video at 24fps with synchronized 48,000 Hz audio, responding to text actions and per-frame camera input. Its WorldPrompt input declares physics rules, character attributes, and camera angles, and it supports multi-person sessions over LiveKit. It also has one eyebrow-raising property: sessions have no preset length. Set against Genie 4's press-account two minutes, that is a natural contrast. Runway's access is similarly restricted, though: approval through a contact form, no public API.
Meanwhile, the other end of the track is consolidating. On September 28, AMD announced an all-stock acquisition of World Labs, Fei-Fei Li's spatial intelligence company whose core product, Marble, generates interactive 3D scenes, for roughly 8.2 billion dollars; once the deal closes, Li joins AMD as chief scientist, reporting directly to Lisa Su. World Labs was founded in April 2024 in San Francisco with roughly 70 people, and its co-founders include Ben Mildenhall, the core inventor of NeRF. The Wall Street Journal reported on October 11 that Amazon is in talks to acquire the world model startup Decart at a valuation of about 7 billion dollars, after Anthropic explored a roughly 6 billion dollar deal and walked away in September; Decart, founded in 2023 by three Israeli engineers, develops the Lucy and Oasis world models and has raised over 400 million dollars. And Rivian acquired Third Dimension, pulling simulation in-house for autonomous driving testing. World models are converging from a startup race into acquisition targets for giants.
In one sentence: world models split between real-time steerable video, the Genie and Runway line, and exportable 3D geometry, the World Labs Marble line, and Genie 4 is the heaviest weight yet thrown onto the first side.
Who Gets Access, and Where the Edges Are
Access comes in three steps. At launch, Genie 4 is open only to trusted testers. The official blog has opened a waitlist for Genie 4 within Project Genie, without a specific date for broader availability. Within weeks, it reaches Google AI Ultra subscribers and early partners on the Gemini API and Google AI Studio, per TechCrunch. There is no public pricing at launch; Ars Technica notes that Google AI Ultra subscriptions start at 100 dollars a month.
The official list of trusted tester backgrounds spans film, journalism, education, gaming, and engineering: filmmakers run multiple storyboard takes in real time; news photographers turn a single archival photo into an environment you can walk into; game developers validate prototypes before writing engine code. Diego Rivas told media that the biggest value of the real-time loop sits at the ideation and storyboarding stage, where creators can test ideas without breaking flow. One developer's description, relayed from X, is "being on a film set": recording multiple takes inside a running simulation, adjusting camera, lighting, and sound on the fly.
On safety, every video and audio output Genie 4 produces carries an invisible SynthID watermark; internal safety evaluations ran before the preview launch, and Google says it applies guardrails consistent with its other models.
The edges deserve to be quoted from the source too: visual details and object permanence still drift over long sessions, repeated complex prompt edits, or scene navigation; the accuracy of complex physical interactions, real-time audio, in-world text rendering, spatial precision, and long-term consistency are all named as active areas of research. And the unpublished list matters just as much: benchmark scores, resolution, frame rate, parameter count, audio language support, audio sample rate, and pricing have all gone undisclosed; do not trust any spec number relayed secondhand. One caution: the details a model invents beyond an archival photo's frame are not historical fact.
Then there is the industry question that will not go away. Bloomberg's headline was that Google upgraded the AI world model that once shook gaming stocks: when the Project Genie prototype landed in January 2026, shares of Unity, Roblox, and Take-Two fell hard. Asked about the impact on gaming, Diego Rivas said the company is working with partners to understand how the technology fits their workflows and where its boundaries are. The question remains open.
Timing sharpens that question. Google reported its third-quarter results the day after the launch, with 2026 capital expenditure guidance revised upward twice to 195-205 billion dollars and free cash flow turning negative in the second quarter. Genie 4 lands at precisely the moment Google's AI investment narrative is under the most scrutiny.
Closing
The real news of the Genie 4 launch is not that the graphics got better. It is that two things happened at once: an AI world gained a native voice, and a language model entered the world model's loop as its director. As for the two minutes, by the TechCrunch and The Verge accounts it is double the previous generation; by Google's more careful wording, this is still a research preview. From the prototype's roughly 60 seconds to the press-account two minutes, the trajectory looks like what video generation went through: short enough to laugh at, rising fast enough to be alarming. The next station worth watching is the subscriber version arriving within weeks, and Runway's no-preset-session-length answer when the two models meet head on.
Related reading