What Is Genie 4? Google DeepMind's Interactive World Model, Explained: What It Actually Does, What DeepMind Admits It Can't, Who Gets It and When, and How It Compares to Genie 3, Marble, Runway, and Odyssey

What a world model is, and what Genie 4 adds

A video generator like Veo renders a clip from a prompt and hands you a file. A world model keeps rendering: it produces the next frame from the last one plus your input, so you can walk, turn, and intervene and the scene responds. Genie 3, announced August 5, 2025, did this at 720p and 24 frames a second with worlds that stayed coherent 'for a few minutes' and a visual memory reaching 'as far back as one minute ago'; it became the engine of Project Genie, a Google Labs experiment for AI Ultra subscribers with generations capped at 60 seconds. Genie 4, per the DeepMind blog by Raia Hadsell, Diego Rivas, and Hang Qi, 'unlocks unprecedented realism and improves real-time interactive simulation.' In its own words, the new pieces:

• Sound, generated with the world. 'Genie 4 is natively multimodal, enabling audio-like music, ambient noises, and interactive dialogue to be simultaneously generated with the world.' And: 'The true frontier, however, is human speech. Genie 4 introduces interactive dialogue that adapts on the fly.' Every sound in the demos 'are produced directly by Genie 4.'
• People and physics. '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.' It 'simulates real-world physics scenarios, such as aerodynamics, collisions, and more.'
• Gemini underneath. '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.'
• An agent in the loop. 'Gemini 3.8 Flash serves as a guiding agent in the loop: as Genie 4 generates a world in real-time, 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.' Between your prompts, 'Gemini also monitors and steers the scene to keep the world progressing cohesively.'
• Steering by text. 'Use natural language to modify a simulation in real-time — like changing the weather, moving the camera, or introducing a new person or object — and Gemini will contextually adapt your prompt to match the surrounding scene.' Shlomi Fruchter, who leads the model, put the contrast plainly on X: 'One of the limitations of Genie 3 was that the worlds were too static.'
• Image to world. The demo prompts begin 'Use the reference image as the exact first frame and visual basis for the scene,' and testers are quoted saying 'a single 2D archival photograph can become a grounded simulation of that environment in seconds.'
• Robots. DeepMind is pairing it with Gemini 4 Argon, which 'breaks down a user-instructed task into steps for Genie to render,' and shows Apptronik's Apollo humanoid playing checkers in a simulation.

What DeepMind says it can't do yet

Quoted whole, because launch posts rarely include a paragraph like this: 'While it can maintain coherent environments over short sessions, 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. Complex physical interactions, real-time audio accuracy, rendering legible text, spatial accuracy, and long-term consistency are active areas of ongoing research.' It calls Genie 4 'an early, experimental model.' In interviews, TechCrunch reported the team 'has not solved accurate audio or legible in-world text,' and Rivas told Bloomberg, 'We're still in the research journey. We're not necessarily in the end product journey.' The number you'll see everywhere, sessions 'up to two minutes, double Genie 3,' comes from Bloomberg and TechCrunch, not from DeepMind's text, which gives no resolution, frame rate, latency, session length, or benchmark at all; Hang Qi told Bloomberg that reaching the limit is no guarantee a scene holds together. Hacker News commenters watching the clips noted a castle wall and catapult vanishing between frames. All of that is consistent with the blog's own caveat. The demo overlays also carry a footer: 'Results for illustrative purposes. Sequences shortened.'

What the demo frames actually show

We matched the frames in this post against DeepMind's clips. The train station is a 25-second demo recording: the blog prompt is 'a crowded rural Japanese train station in the 1980s,' with 'the only railway inspector in a dark navy uniform' standing near the camera, and the text box shows the user typing 'The inspector warns,' then 'He announces the next departure,' then 'An elderly woman arrives with…,' each one a mid-session event the inspector acts and speaks. The Antarctica balloon is a 15-second demo under the 'shaping world simulation in real-time' section; its Reasoning Steps panel reads, verbatim, 'Synthesizing the user's hot air balloon flight over Antarctica from the initial reference image and World State. Establishing initial 3-plane spatial depth and environmental kinetics while keeping the burner valve and vent line controls intact for future player interaction,' and later 'Event 2 is achieved as the balloon floats forward.' That is the Gemini agent's plan made visible, which matches the blog's description of it breaking goals into steps, though neither 'Reasoning Steps' nor 'World State' is named as a product feature anywhere in the text. The hummingbird in the glass sphere is the blog's header art, with no prompt attached. None of the three is in the 147-second launch video on X, which runs through a crowned crane, a Starry Night village, a 1995 Hong Kong motorcycle ride, dragon racing, a lab robot arm, a backstage opera, a space station, horses, a triceratops, a drift car, and a diver, cut with footage of testers at monitors.

Who gets it, when, and what it costs

Three tiers of access, in DeepMind's wording. Now: 'a research preview for trusted testers,' drawn from 'creative industries, journalism, education, gaming, and engineering,' with the stated goal of 'gathering feedback before we release it to consumers and developers.' Next: 'starting with Google AI Ultra subscribers in Project Genie and early access partners of the Gemini API and Google AI Studio,' which Logan Kilpatrick framed as 'coming soon to the API for the first time.' Later: nothing stated. Ars Technica's read is that Google 'may expand access to AI Pro and AI Plus plans eventually' and that there is 'no specific timeline yet'; TechCrunch says 'in the coming weeks.' The waitlist is a Google Form titled 'Apply for Genie 4 access in Project Genie.' Project Genie itself remains an AI Ultra product, 18 and over, now in more than 140 countries; Google's plans page currently lists AI Ultra in 5× and 20× tiers, with the 5× tier shown at $99.99 a month and Ultra 'billed monthly only.' Reports that AI Ultra costs $249.99 are out of date. We have heard from sources that a public launch is planned by November 4; we could not confirm that with Google or find it in any published report, and Google's only word is 'coming soon,' so treat it as a rumor until the Project Genie page changes.

The launch by the numbers

Metrics are from X at 6:15 am Pacific on October 28, read through the fxtwitter API. The brief we received put the launch post at 2.7 million views and 20,859 likes, which checks out. Press was uniform and uncritical on day one (TechCrunch, Ars Technica, Bloomberg, Quartz, The Decoder, Android Authority), with no hands-on reviews, since no outlet has access. The sharpest line came from Rivas himself, on why real time matters at all: 'keeping a state of flow in the creativity space.' The stock angle that defined Project Genie's January launch, when Unity fell more than 20 percent in a day, did not repeat; no market reaction to Genie 4 was reported.

Genie 4 vs Genie 3, Marble, Runway, Odyssey, and the rest

Two readings of that table. Genie 4 is the only system claiming native speech in a live world, and the only one wiring a reasoning model into the loop; those are the real differentiators. It is also closed, unpriced, and undated, in a field where Marble sells you a $20 subscription today, NVIDIA ships weights, and Runway and Odyssey are weeks ahead on public previews. The Hacker News reaction ('Yawn, another closed world model') is the developer side of that. For the consumer the question is simpler: Veo 3.1, our #1 video generator, gives you a finished clip today; Genie 4 will give you a place to stand in, on the most expensive Google plan, at a date Google hasn't said.

Our read

The demos are the best interactive-world footage anyone has published, and DeepMind earns credit for publishing its limits in the same post. The thing to hold onto is the sentence from Rivas: still a research journey, not a product journey. Genie 3 never left Project Genie's sandbox; Genie 4's path out is a waitlist, an AI Ultra tier, and an API 'coming soon.' If you make video for a living, the real-time steering and the agent planning are the features to watch, because they are what turns a world model from a toy into a pre-vis tool, and the testers DeepMind quotes are saying exactly that. If you make games, Unity's verdict on Project Genie, 'unsuitable on its own for games requiring consistent, repeatable player experiences,' still stands until the drift DeepMind describes is gone. And if you want to try it: join the waitlist, keep an AI Ultra subscription warm if you already have one, and don't buy one on a rumor. We'll update this post when Project Genie's page changes, when the API gets a price, and when the first hands-on reports arrive. Related: AI video generators ranking, Gemini 4 Argon is out: how to get it, and what Nano Banana 2.1 changed.