TL;DR
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Google DeepMind introduced Genie 4 on October 27, 2026, describing it as a real-time model for generating interactive worlds with video and audio. The demonstrations are company-selected, public access has not opened, and DeepMind says scenes can drift after a few minutes of interaction.
Google DeepMind announced Genie 4 on October 27, 2026, describing it as a model that generates interactive video worlds with sound as users move through scenes and issue text instructions. The announcement marks a new step in the company’s world-model research, but the system is not publicly available, and DeepMind says objects and details can drift during sessions lasting a few minutes.
DeepMind says Genie 4 generates the next frames of a scene in response to keyboard and mouse controls, as well as written prompts. A user could, for example, ask for a change in weather or request a new object while the scene is running. The company says a Gemini model helps interpret instructions, break goals into steps and monitor the generated scene. These descriptions come from DeepMind; the announcement includes no independent benchmark results.
The company’s demonstrations show environments with ambient sound, music and dialogue generated alongside the visuals. DeepMind calls Genie 4 its first real-time world model to generate audio natively. The article’s source material notes that audio generation is not unique across the industry, citing Runway’s GWM Worlds 2 as another system with audio. DeepMind also describes improved physics and more natural-looking people, but it did not provide quantitative measures for those claims.
Access is currently restricted to trusted testers. DeepMind says the system is expected to come to Google AI Ultra subscribers through Project Genie and to early Gemini API partners. The source identifies the Ultra subscription as costing $200 per month, but gives no date for wider access and no API pricing. TechCrunch and Bloomberg, as cited by heise, have reported demonstrations lasting up to about two minutes; that duration does not appear in DeepMind’s own description, which says interaction can continue for a few minutes.
World models · Research guide · October 2026
A Guide to Google’s Genie 4
Google DeepMind describes Genie 4 as a real-time model for generating interactive worlds with video and audio. Users can move through scenes and request changes while the model predicts what happens next.
01 / How it works
A scene that responds as you explore
Genie 4 predicts the next frames from the scene so far and the user’s actions. DeepMind says a Gemini model interprets goals, breaks them into steps and monitors the generated world.
Move through the scene
Keyboard and mouse input guide movement as the environment updates in response.
Request a new detail
Users can ask for changes during a session, such as different weather or a newly added object.
Sound with the visuals
Demonstrations include ambient sound, music and dialogue generated alongside video.
What is still unmeasured?
DeepMind describes improved physics and more natural-looking people, but the announcement provides no independent benchmark results or quantitative measures. The demonstrations were selected by the company.
02 / Where it could fit
Promising uses, with evidence still early
The company and its testers describe creative and research applications. Those examples suggest possible uses; they do not establish production performance or measured time savings.
Film & game prototypes
Testers have used Genie 4 for film previsualization and game prototyping, according to DeepMind.
Simulated task planning
One demonstration shows Gemini 4 Argon planning steps while Genie 4 simulates Apptronik’s Apollo playing checkers. No physical robot took part.
SynthID watermark
DeepMind says video and audio outputs carry an invisible SynthID watermark. Detection after editing or re-encoding is not explained in the available material.
03 / The research line
From simple scenes to real-time worlds
Each generation has expanded what the model can generate or how people can interact with it, according to Google DeepMind.
Genie
Generated simple 2D worlds. The research paper described an 11-billion-parameter model trained on unlabeled internet videos.
Genie 2
Generated 3D worlds from one image, with consistency up to a minute; real-time operation used a reduced-quality version.
Genie 3
DeepMind’s first real-time world model, described at 720p and 24 frames per second, with scenes lasting a few minutes.
Genie 4
Adds native audio generation and interactive text changes, according to DeepMind.
04 / A different kind of world
Prediction instead of a hand-built map
A conventional game engine renders views from a structured environment. A world model predicts what should appear next from previous frames and user input.
Map → rendered view
A structured world defines objects and layout; the engine renders the view from that stored representation.
Frames → next scene
The model predicts the next view from prior frames and actions. This can create varied scenes without a hand-built map, while allowing details to drift.
05 / What remains open
Limits, access and unanswered questions
The source material describes a research and demonstration release. Broader testing will be needed to judge reliability, output quality and practical cost.
Evidence to watch
- No independent benchmark results are included in the cited announcement material.
- Scenes may drift over time; DeepMind describes continuous interaction lasting a few minutes.
- TechCrunch and Bloomberg, as cited by heise, report demonstrations up to about two minutes. That figure is not in DeepMind’s own description.
- Simulated physics transfer to real-world robotics has not been established by this example.
Access still pending
- Currently limited to trusted testers.
- Planned for Google AI Ultra subscribers through Project Genie and early Gemini API partners.
- The source identifies Ultra at $200 per month.
- No launch date for broader access or Gemini API pricing is provided.
What would make the next update meaningful?
Expanded access could allow independent testing of scene coherence over longer sessions, audio and physics quality, and how well text controls handle varied prompts. Practical value will also depend on availability, performance and a workable cost.
06 / Key questions
Genie 4, in brief
What is Google DeepMind’s Genie 4?
A world model that DeepMind says generates interactive scenes frame by frame in response to movement and text instructions, with audio generated alongside visuals.
Can the public use it now?
No public release is described. Access is limited to trusted testers, with planned availability through Project Genie for Ultra subscribers and for early Gemini API partners. No date is given.
How long can a scene run?
DeepMind says a few minutes of continuous interaction and warns that objects and details can drift. Separate reports cited by heise say demonstrations lasted up to about two minutes.
Has it trained a real robot?
The cited Apollo checkers example is a simulation; no physical robot took part. The material does not establish Genie 4 deployment for training physical robots.
Interactive Worlds for Games and Robotics
If the capabilities described by DeepMind hold up outside its selected demonstrations, Genie 4 could give game developers and filmmakers a way to sketch scenes they can explore and revise in real time. DeepMind says testers have used the system for film previsualization and game prototypes. Those examples are tester feedback reported by the company, not evidence of established production use or measured time savings.
The robotics demonstration points to a separate research use: generating simulated environments for planning and training. DeepMind describes a task in which Gemini 4 Argon planned steps and Genie 4 simulated Apptronik’s Apollo humanoid playing checkers. According to the source material, no physical robot took part. Waymo already uses the predecessor, Genie 3, in its driving simulator, but that does not establish that Genie 4 is being used for real-world robot training.
Generating audio with a scene may also make synthetic clips more convincing. DeepMind says Genie 4 video and audio outputs carry an invisible SynthID watermark. The available material does not explain how reliably that watermark can be detected after editing or re-encoding, or how it will be exposed to viewers.
Products Worth Considering
Logitech MX Master 4 for Mac, Ergonomic Bluetooth Mouse with Haptic Feedback, Ultra-Fast Scrolling, USB-C Charging, for MacOS, iPadOS, Windo
Logitech MX Mechanical Wireless Illuminated Keyboard, Tactile Quiet, Graphite
Logitech MX Keys S Wireless Keyboard, Graphite
As an affiliate, we earn on qualifying purchases.
From 2D Scenes to Real-Time Worlds
DeepMind’s Genie models have developed from generating simple environments toward interactive, real-time scenes. The first Genie, introduced in February 2024, generated simple 2D worlds; its research paper described an 11-billion-parameter model trained on unlabeled internet videos. Genie 2, announced in December 2024, generated 3D worlds from a single image and could keep them consistent for up to a minute, although real-time operation used a reduced-quality version.
In August 2025, Genie 3 became DeepMind’s first world model to run in real time. The company described it as operating at 720p and 24 frames per second, with scenes remaining consistent for a few minutes. Genie 4 builds on that line of work by adding native audio generation and interactive text changes, according to DeepMind.
A world model differs from a conventional game engine, which stores a structured map and renders a view from it. Genie 4 instead predicts what should appear next based on previous frames and user input. That approach can produce varied scenes without a hand-built map, but the model’s reliance on prediction also helps explain why objects may change or drift as the interaction continues.
“Details and objects “can shift or drift over time.””
— Google DeepMind
Limits and Access Still Open
DeepMind has not published benchmark results in the source material, so claims about realism, physics and the quality of its audio cannot yet be compared against independent measurements. The demonstrations were selected by the company, and the article reports no independent evaluation of how well the scenes respond to varied instructions or remain consistent over longer sessions.
Public access details remain incomplete. Google has not provided a launch date for Project Genie access or early Gemini API availability, and no API price is stated. The reported duration of up to about two minutes comes from TechCrunch and Bloomberg as cited by heise; DeepMind’s own wording is “a few minutes.” It is also unclear how much performance or output quality may differ between the demonstrations and access given to testers.
The robotics example is a simulation, not a report of a real humanoid completing the task. The source material does not establish that Genie 4 has been deployed for training physical robots, or provide evidence about how faithfully its simulated physics transfer to real-world conditions.
Waiting for Broader Testing
The next concrete milestone is access for Google AI Ultra subscribers through Project Genie and for early partners using the Gemini API. Google has not announced when either group will receive access, how many testers will take part or what the API will cost. Independent testing after access expands should help establish how long scenes remain coherent and how well audio, physics and text controls perform across different prompts.
For now, Genie 4 is best understood as a research and demonstration release with potential uses described by DeepMind and its testers. Its practical value for production workflows and robotics will depend on results beyond selected clips, including how reliably it handles longer sessions and whether developers can access it at a workable cost.
Key Questions
What is Google DeepMind’s Genie 4?
Genie 4 is a world model that DeepMind says generates interactive scenes frame by frame in response to user movement and text instructions. The company says it can generate audio alongside the visuals.
Can the public use Genie 4 now?
No public release is described in the source material. DeepMind says access is limited to trusted testers, with availability planned for Google AI Ultra subscribers through Project Genie and early Gemini API partners. No launch date is given.
How long can a Genie 4 scene run?
DeepMind says Genie 4 supports a few minutes of continuous interaction and warns that objects and details can drift over time. TechCrunch and Bloomberg, as cited by heise, report demonstrations of up to about two minutes; that figure is not in DeepMind’s own text.
Has Genie 4 been used to train a real robot?
The robotics example in the source material is a simulation involving Apptronik’s Apollo humanoid playing checkers; it says no physical robot was involved. Waymo uses the earlier Genie 3 in its driving simulator, but that is separate from evidence of Genie 4 training real robots.
Source: Thorsten Meyer AI
