Google DeepMind unveils Genie 4, a real-time world model with native audio, in research preview
Google DeepMind announced Genie 4 on October 27, 2026, presenting it as a new generation of its real-time interactive world model and the first in the Genie line to generate sound natively alongside the image. Each simulated world now arrives with its own ambient audio, music and interactive dialogue, and users can reshape a running scene with natural-language prompts, changing the weather, moving the camera or adding people and objects while the world keeps evolving between inputs.
The model is not yet available to the public. Google DeepMind has released Genie 4 as a research preview to a limited group of trusted testers and says it is "coming soon" to Google AI Ultra subscribers in Project Genie, the experimental Google Labs app that has so far run on Genie 3, and to early-access partners of the Gemini API and Google AI Studio. A waitlist for Project Genie is open. Google has not published pricing, a release date, supported countries or API documentation for the new model.
The announcement is also notable for what it leaves out. Unlike the Genie 3 release, which came with headline figures for resolution and frame rate, the Genie 4 post contains no technical specifications, no benchmark results and no direct numerical comparison with its predecessor. What it does describe in detail is a change in architecture: Genie 4 is built on top of the Gemini model family, with Gemini Omni grounding appearance and behavior and other Gemini models handling the reasoning that keeps a simulation coherent over time.
··········
WHAT GENIE 4 DOES IN A RUNNING SIMULATION.
A world model in Google DeepMind's sense is a system that generates an environment frame by frame in response to a user's actions, rather than rendering a pre-built 3D scene. In the demonstrations published with the announcement, users moved through worlds with standard keyboard controls for navigation and camera, and used text prompts to steer what happens next. Google DeepMind says Gemini adapts each prompt to fit the scene already on screen, so an instruction to change the weather or introduce a character is integrated into the existing world instead of producing a cut to a new one.
The published demos illustrate the range the company is aiming for. One shows a saxophone player on a floating platform in a misty lake, with a camera that moves freely around the performer and the ambient sound of the water. Another follows hot air balloons over Antarctica from an aerial view before switching to a first-person view from inside a basket. A third renders a long, billowing pink satin river winding through a wildflower meadow, a deliberately unreal scene meant to show that the model is not limited to photorealistic reconstruction.
Google DeepMind also says Genie 4 simulates real-world physics such as aerodynamics and collisions, and that characters move naturally, react to their surroundings and can be steered in real time. The physics examples include a macro shot of a maple seed falling into grass, a glossy watermelon on a stone plinth with droplets caught in backlight, and a crowded 1980s rural Japanese train station with an inspector and a long red train in the foreground. These are vendor-selected demonstrations: the company has not released a quantitative evaluation of physical accuracy, so they show what the model can produce at its best rather than how consistently it does so.
··········
NATIVE AUDIO AND INTERACTIVE DIALOGUE.
The clearest functional change from Genie 3 is sound. Google DeepMind describes Genie 4 as its first real-time world model to generate audio natively, which means the soundscape is produced by the same system that produces the frames, on the fly, instead of being added afterwards by a separate tool. Each world carries the ambient sound of its setting, and characters can hold conversations that adapt to user inputs and to events in the scene.
The audio demonstrations were chosen to show sound tied to motion and place. A first-person ride on a black dragon through a mountain race is accompanied by wingbeats, wind and a volcanic rumble. A shot styled as a 1995 Hong Kong romantic drama follows a motorcycle crossing a bridge at dawn. A stained-glass fantasy landscape with a river, flowers and mountains can be explored in 3D with its own ambient track.
The company is candid that this part of the system is immature. Among the open research problems it lists is "real-time audio accuracy", alongside complex physical interactions and the rendering of legible text. For users, that means the audio should be treated as an integral but still imperfect layer of the simulation, especially where dialogue or sound needs to match on-screen action precisely.
··········
HOW GEMINI MODELS STEER THE SIMULATION.
The most significant technical shift in Genie 4 is the role of Gemini. Google DeepMind says Gemini Omni anchors simulations in real-world appearance and behavior, while other Gemini models provide the reasoning that keeps a world coherent and continuous. The post does not explain how these components are connected, whether they run as separate models or share weights, or how much compute a session requires.
For complex events, the company uses Gemini 3.8 Flash as a guiding agent during real-time generation. The agent watches the scene, breaks a goal into steps, checks incoming frames to confirm that each step has succeeded, and revises its plan when the world shifts or a user intervenes. The demonstrations for this mode include a UR5 robot arm dispensing liquid into a lab plate, a first-person motorcycle rider stopping at a red light beside a beach, and a crowned crane walking across snow at dusk.
A second example pairs Genie 4 with Gemini 4 Argon, Google's current frontier model, and Gemini Robotics. In that setup, Argon splits a user's task into steps that are fed to Genie 4 as text captions during rendering. The published demo shows Apptronik's Apollo humanoid robot playing checkers, with the game broken down into real-time steps. For readers following Google's model line, the role of Argon here connects to the capabilities described at its launch in our coverage of the Gemini 4 Argon global release.
Google DeepMind frames this combination as the start of what it calls hybrid world models, in which the structured reasoning of language and code models is combined with pixel-level sensory simulation. In the early version shown, Gemini maps the structure and plans the narrative while Genie renders how the world looks, sounds and responds. The authors say they believe this approach could produce "a new class of world models that are far more coherent, controllable, and grounded". That is a stated research direction, not a measured result.
··········
GENIE 4 COMPARED WITH GENIE 3 AND PROJECT GENIE.
Because Google DeepMind has not published specifications for Genie 4, the only reliable comparison is between what each release documents. The table below sets out the figures and features Google has stated officially for Genie 3, for the current Project Genie prototype, and for Genie 4.
........
Item | Genie 3 | Project Genie (Genie 3) | Genie 4 |
Resolution | 720p | Not stated | Not stated |
Frame rate | 20-24 fps | Not stated | Not stated |
Consistency | Several minutes | 60-second generations | Short sessions |
Native audio | No | No | Yes, incl. dialogue |
Promptable events | Yes | Not included | Yes |
Gemini integration | Not described | Gemini, Nano Banana Pro | Gemini Omni, 3.8 Flash, 4 Argon |
Public access | Via Project Genie | Google AI Ultra, 18+ | Trusted testers only |
........
The comparison shows why the absence of numbers matters. Genie 3 was documented at 720p and 20 to 24 frames per second, with environments that stay largely consistent for several minutes and a memory of specific interactions lasting up to about a minute. Project Genie, the consumer prototype built on it, limits generations to 60 seconds and, at its launch in January, did not include the promptable world events that Genie 3 had demonstrated. Genie 4 brings promptable events, physics and audio together in one system, but Google has not said whether it matches or exceeds Genie 3's resolution, frame rate or session length.
The interactive world-model field has become more crowded since Genie 3. Data Studios has covered open-weight alternatives such as Tencent's Hy-Nomas interactive world model, which can be downloaded and run independently, while Genie 4 remains a closed model accessible only through Google's own products.
··········
AVAILABILITY, ACCESS AND WHAT USERS CAN DO NOW.
For now, Genie 4 can only be used by a limited group of trusted testers chosen by Google. The company's official wording is that the model "has been released as a research preview to trusted testers, and is coming soon to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and Google AI Studio". No date has been given for either step, and the post does not say whether Google AI Pro or free Gemini users will get access.
Project Genie is the route most consumers will see first. Google's subscriptions page lists Project Genie as available only with a Google AI Ultra plan and still describes it as running on Genie 3. Ultra is listed from $99.99 per month, with a higher tier at $199.99 per month that offers 20 times the usage limits of AI Pro. When Project Genie launched in January 2026 it was limited to Ultra subscribers in the United States aged 18 and over, and was not available to Google AI Ultra for Business users. In May, Google said the app was gradually rolling out to eligible Ultra subscribers worldwide aged 18 and over, and added Street View grounding for locations in the United States. Google has not said whether the move to Genie 4 will change these eligibility rules or the 60-second generation limit.
Developers have no general access path yet. Genie 4 will reach the Gemini API and Google AI Studio only through an early-access program, and Google has published no model identifier, pricing, rate limits or documentation. Teams that want to evaluate it for agent training, simulation or prototyping work will need to apply for early access once Google opens it; the existing Gemini API changelog contains no Genie 4 entry.
Users who want to try the model when it reaches Project Genie can join the waitlist linked from the Google DeepMind announcement. Anyone planning to subscribe to Ultra specifically for Genie 4 should note that Google has given no rollout date, so the subscription currently buys access to the Genie 3 version of the app.
··········
EARLY USES REPORTED BY TRUSTED TESTERS.
Google DeepMind describes four groups of early users. Filmmakers and creative directors are using Genie 4 as a "draft mode" for previsualization, running multiple takes of a scene and adjusting camera, lighting, dialogue and action live before committing to a shoot. A second group is turning single archival photographs into explorable environments and simulating historical scenarios and events such as floods or wildfires, an application the company links to journalism and education.
A third group works on interactive experience design, using Genie 4 to move from an idea to an explorable prototype and to test visual styles and mechanics through prompts before writing any game-engine code. The fourth consists of enterprise partners in simulation, who are evaluating what the company calls fast "zero-to-one" scenario generation and identifying gaps the model still needs to close.
These descriptions come from Google, and no tester has been named or quoted with results. They do indicate where the company sees near-term value: tasks where a fast, approximate, controllable simulation is more useful than a precise one, and where drift over long sessions matters less than speed of iteration.
··········
SAFETY MEASURES AND STATED LIMITATIONS.
Google DeepMind says Genie 4 includes safety guardrails aligned with its policies for other models and went through internal safety evaluations aimed at preventing harmful or malicious content before the preview. The staged rollout to trusted testers is presented as part of that approach. All video and audio generated by Genie 4 will carry SynthID watermarks, covering both modalities, which matters for a model capable of producing realistic footage with matching sound and speech.
The company's list of limitations is specific. It says Genie 4 "remains an early, evolving area of research" and that, while it can maintain coherent environments over short sessions, visual details and object permanence can shift or drift over time. Drift is more likely when a simulation runs longer, after multiple complex prompt edits, or while navigating through a scene. Complex physical interactions, real-time audio accuracy, legible text rendering, spatial accuracy and long-term consistency are all named as active research areas.
··········
TECHNICAL SYNTHESIS.
What is established is that Google DeepMind has a new world model that generates interactive scenes with native audio and dialogue, accepts natural-language edits to a running simulation, and is steered by Gemini models acting as planners and checkers. The company has demonstrated it with selected examples covering physics, character motion, sound, and robotics planning with Gemini 4 Argon and the Apollo humanoid, and has committed to SynthID watermarking for every output.
What remains uncertain is almost everything that would allow an independent assessment: resolution, frame rate, latency, maximum session length, compute requirements, how Gemini Omni and Genie 4 are connected, and how the model performs against Genie 3 on any measured task. Access is limited to trusted testers, with Project Genie and the Gemini API both marked only as coming soon. Until Google publishes specifications or opens the model more widely, Genie 4 is best read as a significant change in design, with Gemini reasoning and native audio added to the Genie line, rather than a product whose capabilities can yet be verified.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
