Google Genie 4: What DeepMind’s New World Model Can Do

Google DeepMind has unveiled Google Genie 4, a new AI “world model” that turns a prompt into an interactive scene you can explore in real time. Announced on 27 October 2026, Genie 4 adds something its predecessors lacked: sound. Wind, waves, music and even spoken dialogue are generated together with the world itself, and you can change the scene while it is running by typing new instructions.

For now, Genie 4 is a research preview limited to trusted testers. Google says it is coming soon to Google AI Ultra subscribers through Project Genie, and to early-access partners of the Gemini API and Google AI Studio. This guide explains what Genie 4 is, what is new, how it works with Gemini, where it still falls short, and when you may be able to try it.

Google Genie 4 at a glance

• Announced: 27 October 2026 by Google DeepMind
• What it is: a real-time, interactive world model with native audio
• How you control it: keyboard keys for movement and camera, text prompts to change the scene
• Powered by: Gemini models, including Gemini Omni for realism and Gemini 3.8 Flash as a guiding agent
• Status: research preview for trusted testers
• Next: Google AI Ultra subscribers in Project Genie, and early-access partners in the Gemini API and Google AI Studio
• Release date and price: not announced
• Safety: all generated video and audio carries a SynthID watermark

What is Google Genie 4?

Genie 4 is a world model: an AI system that generates an environment frame by frame and updates it as you act. A text-to-video tool produces a finished clip and stops. A world model keeps simulating what happens next. You move through the scene with standard keyboard controls and steer events with text prompts, and the model answers with new pictures and sounds in real time.

DeepMind calls Genie 4 natively multimodal. Ambient noise, music and character speech are produced at the same moment as the visuals, rather than being added afterwards. The team argues that understanding and simulating the world is a crucial step towards artificial general intelligence (AGI). It also sees world models as a way to build safe, endlessly varied training grounds for robots and AI agents, as well as new kinds of interactive entertainment.

What is new in Genie 4?

DeepMind groups the upgrade into four main capabilities. Here is what each one means in practice.

Shaping the world while it runs

With Genie 4 you can type instructions to change the weather, move the camera, or introduce a new person or object while the simulation is playing. Gemini adapts each prompt so that it fits the surrounding scene. Between your inputs, Gemini also watches the world and nudges it forward so the story stays coherent. DeepMind’s demonstrations include a saxophone player on a floating platform in a misty lake, a hot air balloon drifting over Antarctica, a crowded Japanese train station in the 1980s, and a landscape made entirely of stained glass.

More realistic physics and people

DeepMind says Genie 4 rests on a stronger grasp of how the physical world behaves, including aerodynamics and collisions. It describes the realism of people as the biggest leap: characters move naturally, react to their surroundings and can be steered live. One demonstration is a macro close-up of a maple seed falling into grass, which tests fine motion and air resistance.

Native audio and live dialogue

Genie 4 is DeepMind’s first real-time world model to generate audio natively. Every world carries the ambient sound of its own setting, such as wind in trees or waves on a beach. The more striking feature is speech: characters can hold conversations that adapt to your inputs and to events in the scene, all in real time.

Gemini as a planner inside the simulation

The fourth capability is reasoning. By pairing Genie 4 with Gemini, DeepMind can simulate complex, multi-step events and test how AI systems cope with different environments before they are used in the real world. The next section explains how that loop works.

How Genie 4 works with Gemini

Genie 4 builds on the Gemini family rather than standing alone. According to DeepMind, Gemini Omni anchors the simulations in real-world appearance and behaviour, while other Gemini models supply the reasoning needed to keep a simulation consistent over time.

To make that work in real time, Gemini 3.8 Flash acts as a guiding agent inside the loop. While Genie 4 generates the world, Gemini keeps observing the scene, breaks a goal into a sequence of steps, and checks incoming frames to confirm that each step succeeded. If the world shifts or the user intervenes, Gemini changes its plan on the spot. The aim is worlds with reliable cause and effect, procedural behaviour and multi-step reasoning.

DeepMind is also testing the combination with its robotics research. In one experiment, Gemini 4 Argon breaks a user’s instruction into steps and feeds them to Genie 4 as text captions while the simulation renders. One example shows a simulated Apptronik Apollo humanoid playing checkers. Another shows a Universal Robots UR5 arm using a pipette in a laboratory scene. Because the captions can change as the scene unfolds, the simulation can adapt as the task moves towards completion.

Looking ahead, DeepMind points to “hybrid world models”. These would combine the structured reasoning of code and language models with the rich, pixel-level detail of a simulator like Genie. In that picture, Gemini maps the structure and narrative timeline behind the scenes, while Genie renders how the world looks, sounds and responds.

Genie 4 vs Genie 3: what has changed?

Genie 3, announced in August 2025, was DeepMind’s earlier real-time world model. Its launch post cited 720p video at 24 frames per second, with visual memory reaching back about a minute. Google then opened a prototype called Project Genie, powered by Genie 3, to Google AI Ultra subscribers in the United States aged 18 and over on 29 January 2026. At that stage, promptable events were not yet part of the prototype.

Genie 4 adds native audio and dialogue, live text steering guided by Gemini, and more lifelike people. Reports from Google’s launch briefing say sessions run for roughly two minutes. DeepMind did not publish resolution, frame rate or memory figures for Genie 4 in its announcement, so a like-for-like comparison of specifications is not possible yet. Independent hands-on testing of the earlier Project Genie prototype also found visible continuity failures, which is worth remembering when judging any demo reel.

What can Genie 4 be used for?

DeepMind has been working with trusted testers in creative industries, journalism, education, gaming and engineering. They point to four early uses:

• Creative direction: filmmakers can use Genie 4 as a quick draft mode for pre-visualisation, stepping into a live scene and running several “takes” while changing camera angles, lighting, dialogue and action on the fly.
• Historical scenarios and immersive photographs: photojournalists are testing whether a single archival photograph can become a grounded simulation of that place within seconds, and whether short events such as a flash flood or wildfire can be visualised as they unfold.
• Interactive concept design: game developers are using text prompts to go from an idea to an explorable prototype, testing visual style, materials and mechanics before writing any engine code.
• Simulation: enterprise partners are evaluating Genie 4 for fast “zero-to-one” scenario generation and are helping DeepMind find the gaps that still need work.

The longer-term prize is robotics and agents. A system that can generate endless, varied and believable situations gives researchers a cheap place to test how a robot or AI agent behaves before it meets the real world. That is why DeepMind links Genie so closely to its Gemini Robotics work and to its AGI goals. Agents themselves are advancing quickly too, as our report on OpenAI’s Dots agents shows.

Limitations and open questions

DeepMind is open that Genie 4 is early, experimental research. It lists several weak spots:

• Visual details and object permanence can drift, especially in longer sessions, after several complex prompt edits, or when you move around the scene.
• Complex physical interactions are still hard to simulate accurately.
• Real-time audio accuracy and legible on-screen text remain open problems.
• Spatial accuracy and long-term consistency are still active areas of research.

There are also gaps in what has been disclosed. DeepMind has not published the model’s size or benchmark results, and it has not said how Genie 4 runs on consumer hardware. Because the demonstrations were chosen by the company, treat them as best-case examples until independent testers can try the model themselves.

Safety, guardrails and SynthID

Google says Genie 4 has strict safety guardrails in line with its policies for other models, and that it ran internal safety evaluations to prevent harmful or malicious content before the preview. Releasing the model first to trusted testers is part of the same careful approach. All video and audio generated by Genie 4 includes SynthID, Google’s imperceptible watermark, which helps identify AI-generated content in both pictures and sound.

Genie 4 release date, access and price

Genie 4 is not open to the public yet. Here is what is known:

• Now: a research preview for trusted testers.
• Coming soon: Google AI Ultra subscribers through Project Genie, and early-access partners in the Gemini API and Google AI Studio. Press reports talk about “the coming weeks”, but Google has not given a date.
• Waitlist: Google has opened a waitlist for Genie 4 in Project Genie.
• Price: no separate price has been announced. Access is tied to the Google AI Ultra subscription, which some reports put at roughly $100 a month, so check Google’s own pricing page.
• Countries: Google has not listed supported countries for Genie 4. Project Genie itself began in the United States and has since been rolling out more widely to eligible adult subscribers.

How Genie 4 compares with other world models

Google is not alone in this race. Runway released its second world model, GWM Worlds 2, a few months ago, and it also reacts to user input and generates audio, according to reports. DeepMind is also developing SIMA 2, an agent that plays, reasons and learns in virtual 3D worlds, which shows how closely the company ties simulated environments to agent research.

Where Genie 4 stands out is the pairing with Gemini. Using a language and reasoning model as a live director, rather than relying on the world model alone, is meant to keep scenes coherent and give them goals. Whether that advantage holds up outside Google’s own demonstrations is the main question testers will answer in the coming months.

What Genie 4 means for creators, developers and everyday users

For creators, the appeal is speed. A director or designer can explore a scene, try different camera moves and hear how it sounds within minutes, then decide what is worth building properly. The model is a sketchbook, not a replacement for a finished production.

For developers, early access through the Gemini API and Google AI Studio could make world simulation a building block for games, training tools and testing pipelines. Pricing and usage limits will decide how practical that is.

For everyone else, the questions are about trust and ownership. SynthID helps people spot AI-generated content, but it does not settle who owns what a world model produces. Our explainer on who owns AI-generated work shows that purely machine-made output can struggle to earn copyright protection, so creators should keep records of their own creative choices.

One practical point for 3D artists: Genie 4 generates worlds frame by frame that you explore, much like an interactive video. Google has not announced a way to export finished 3D models or game-ready assets from it, so it does not replace modelling software.

Google Genie 4: frequently asked questions

Sources

• Google DeepMind: Genie 4, a new generation of world simulation
• Google DeepMind: Genie model page
• TechCrunch: Genie 4 adds audio, more realistic worlds and real-time steering
• The Next Web: Genie 4 builds AI worlds that talk and change on command
• PPC Land: Genie 4 simulates people who talk back in real time
• Free Press Journal: Genie 4 for real-time AI worlds with built-in audio
• eesel AI: Genie 4, Google DeepMind’s real-time world model, explained
• DataNorth: Genie 4 adds sound and talking characters to playable AI worlds