Genie 4: Google DeepMind's Real-Time AI World Model Explained

Google DeepMind unveiled Genie 4 on October 27, 2026, a world model that generates interactive 3D environments in real time, complete with native audio, physics, and human characters that respond to text or keyboard steering as the scene unfolds. It is live only as a research preview for trusted testers, with a waitlist open for wider access through Project Genie and the Gemini API.

What is Genie 4?

Genie 4 is Google DeepMind's latest "world model" — a system that simulates an interactive, navigable environment rather than simply generating a fixed video clip. According to DeepMind's announcement post, the model is "natively multimodal," which the company says means it generates "music, ambient noises, and interactive dialogue" at the same time as the visuals, rather than adding sound in a separate pass. A person can move through the generated world with standard keyboard controls while also typing natural-language instructions — changing the weather, repositioning the camera, or adding people and objects — and the scene adapts on the fly.

DeepMind frames the release within its broader AGI research agenda, writing that "understanding and simulating the world is a crucial aspect of building AGI." That framing puts Genie 4 alongside other Gemini-family releases the company has shipped this year, including the Gemini 4 Carbon model line, though Genie is a distinct product built for simulation rather than chat or coding. DeepMind published the full announcement, "Genie 4: A new generation of world simulation," on its own blog, credited to Raia Hadsell, Diego Rivas, and Hang Qi.

Genie is now DeepMind's third generation of this research line. The original Genie, published as a research paper, could only generate short, simple 2D platformer-style scenes. Genie 3, announced in August 2025, moved to real-time 3D exploration at 720p and 24 frames per second with no native sound. Genie 4 keeps the real-time, explorable format but adds sound, steerable characters, and the Gemini-guided agent loop described below — the step DeepMind is treating as the bigger leap this time.

How Genie 4 works

Genie 4 is built on top of the Gemini model family rather than standing alone. DeepMind says Gemini Omni "anchors the simulations in real-world appearance and behavior," supplying the visual and physical grounding for what a scene should look and feel like, while other Gemini models handle the reasoning needed to keep a long interactive session coherent — remembering what already happened in the scene and making sure new elements fit what came before.

Sitting inside that loop is Gemini 3.8 Flash, which DeepMind describes as "a guiding agent in the loop." The agent observes the current state of the simulated world, breaks a user's goal into a sequence of smaller steps, checks rendered frames to verify each step actually happened as intended, and revises its plan if the world drifts or the user intervenes mid-scene. This is a meaningfully different architecture from a one-shot video generator: Genie 4 is closer to an agent that continuously directs a simulation than a model that renders a clip and stops.

In practice, that means a user's prompt doesn't just get rendered once. If someone asks the scene to "make it start raining" or "add a second person walking into frame," Gemini 3.8 Flash has to decide what sequence of visual and physical changes would actually produce that outcome, hand instructions to the rendering process, and then check the output against the goal before moving on. DeepMind doesn't disclose how often that verification step catches and corrects an error versus how often drift still slips through, which ties directly into the limitations the company lists elsewhere in the post.

Native audio: the headline feature

DeepMind is pitching audio as Genie 4's signature advance. The company states plainly that "Genie 4 is our first real-time world model to generate audio natively," covering ambient sound and real-time interactive dialogue generated alongside the moving image, not layered on afterward. Previous Genie generations, including Genie 3, were silent — audio synthesis tied to a live, steerable 3D scene is new to this release. DeepMind has not published technical specifics such as sample rate, latency, or how dialogue is kept in sync with lip movement or camera cuts, so those mechanics remain unconfirmed outside the company's own description.

The framing matters because audio is normally the hardest piece to bolt onto a generative video system after the fact — timing has to track a camera that the user is actively moving, and dialogue has to respond to a scene that didn't exist a moment before someone typed a new instruction. By building sound generation into the same real-time loop as the visuals and the physics, DeepMind is arguing that Genie 4's audio is reactive to steering in the same way the picture is, rather than a soundtrack stitched on top. Whether that holds up under sustained, independent testing is one of the first things reviewers with preview access are likely to probe once broader testing begins.

Physics, movement, and human characters

Genie 4 simulates physical interactions such as aerodynamics and collisions within its generated worlds, and DeepMind says human characters in a scene "move naturally, react to their environment," and can themselves be steered in real time rather than just following a scripted animation. The company also highlights "enhanced photorealism and fluid movement" as general visual improvements over earlier Genie releases. DeepMind did not publish benchmark scores, resolution figures, frame rate, or maximum scene duration in the announcement, so none of those numbers can be confirmed from the primary source — treat any such figures you see cited elsewhere as unverified until DeepMind publishes them directly.

The robotics demo: Genie 4 meets Gemini Robotics

Alongside the core release, DeepMind showed a separate experiment pairing Genie 4 with its robotics work. In that demo, a different Gemini model — Gemini 4 Argon — "breaks down a user-instructed task into steps for Genie to render," with those steps interleaved as text captions while the scene plays out. The showcased example has Apptronik's Apollo humanoid robot playing a game of checkers inside a Genie 4 simulation. DeepMind frames this as early evidence that world models like Genie 4 can serve as training and evaluation environments for physical robots and other agents, ahead of or alongside real-world deployment, though the company has not released performance data from the demo. It's worth keeping the two systems straight: Gemini 3.8 Flash is the general-purpose steering agent built into Genie 4 itself, while Gemini 4 Argon is a separate model DeepMind paired with Genie 4 specifically for this robotics experiment — the two names show up interchangeably in some secondary coverage, but DeepMind's post describes them as distinct roles.

Who can access Genie 4, and how to join the waitlist

Genie 4 is not broadly available. DeepMind's own rollout plan, laid out in the announcement, moves through three tiers:

DeepMind points interested users to a waitlist form for Project Genie access, which the company links directly from the announcement post. No price has been disclosed for Genie 4 access itself; DeepMind has not stated whether it will be bundled into existing Google AI Ultra pricing or billed separately once it leaves preview. Developer access, when it opens, will run through the existing Gemini API and Google AI Studio developer platform rather than a standalone Genie account. Given that access runs through Google's own developer accounts, it's worth noting that programs like these authenticate requests the same way most of Google's developer platforms do, via OAuth 2.0-based sign-in.

What Genie 4 is for

DeepMind lists several intended uses in the announcement rather than positioning Genie 4 as a consumer toy:

• Robotics and agent training — generating varied simulated environments for training and testing physical or software agents, as shown in the Apollo checkers demo.
• Creative pre-visualization — DeepMind suggests filmmakers could use Genie 4 to run through multiple "takes" of a scene before committing to a real shoot.
• Turning archival photos into environments — a single still image can be expanded into an explorable simulated space.
• Modeling events — DeepMind cites examples like simulating floods or wildfires, which could support planning or training use cases.
• Game and interactive prototyping — quickly testing interactive scenes or mechanics without building a full engine environment.
• Enterprise "zero-to-one" scenario generation — DeepMind's own phrasing for using Genie 4 to stand up a simulated scenario from nothing for business use cases.

These use cases echo the direction of DeepMind's other applied research efforts, such as its program expanding AI partnerships for scientific research at rival labs — simulation and world modeling are increasingly pitched as infrastructure for training other AI systems, not just as content-generation tools for people.

Limitations and safety guardrails

DeepMind is explicit that Genie 4 has rough edges. The company states that visual details and object permanence "can drift" over long sessions, after many prompt edits, or during extended navigation — meaning an object or detail introduced earlier in a scene may not persist correctly as the simulation continues. DeepMind also lists complex physical interactions, real-time audio accuracy, legible on-screen text, spatial accuracy, and long-term scene consistency as areas it considers still-active research problems, not solved capabilities.

On safety, DeepMind says it applied "strict safety guardrails" consistent with its policies for other models, ran internal safety evaluations before opening the preview, and is deliberately staging the rollout to trusted testers before widening access. Every video and audio clip Genie 4 produces is tagged with SynthID, Google's imperceptible watermarking system for identifying AI-generated media across image, audio, video, and text outputs.

How Genie 4 compares to Genie 3

DeepMind's announcement references Genie 3, the prior generation the company introduced in August 2025, as context for this release but does not publish a detailed side-by-side comparison or benchmark table for Genie 4 itself. That earlier post did disclose concrete specs: Genie 3 rendered at 720p and 24 frames per second, held visual memory of roughly one minute, and kept a scene consistent for a few minutes of real-time navigation. DeepMind has not published equivalent resolution, frame-rate, or duration figures for Genie 4, so a precise numeric comparison isn't possible from public information yet.

The clearest documented change, in other words, is audio: Genie 3 did not generate sound, while Genie 4 is described as DeepMind's first real-time world model built to generate audio natively. DeepMind also cites general improvements in photorealism and movement fluidity for Genie 4 without quantifying them. Readers should treat any specific numeric comparisons circulating elsewhere — claims about exact scene-duration limits or frame rates relative to Genie 3, for instance — as unconfirmed, since DeepMind's own Genie 4 post does not state them.

Frequently asked questions

See the FAQ section below for quick answers on release date, access, pricing, and what makes Genie 4 different from a standard video generator.

What happens next

The near-term path is access, not features: DeepMind says Google AI Ultra subscribers will get into Genie 4 through Project Genie, and developers will be let in through the Gemini API and Google AI Studio, both described as "coming soon" rather than given firm dates. Anyone wanting in early can join the Project Genie waitlist DeepMind links from its announcement. Beyond access, the open questions are the ones DeepMind itself flagged — whether drift in object permanence and physics accuracy improve enough for longer, more reliable sessions, and whether the robotics pairing with models like Gemini 4 Argon moves from a checkers demo into something resembling a standard training pipeline. DeepMind has not given a timeline for general availability or published pricing for Genie 4 once it exits preview.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage