🔍 Read the full analysis: The Promise And Problems Of Google’s Genie 4 AI Worlds on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Google DeepMind presented Genie 4 on October 27, describing a real-time model that creates interactive scenes with synchronized audio and text-directed changes. The demonstrations are company-selected, the model is not yet broadly available, and DeepMind says scenes can drift after a few minutes; independent benchmark results and a release date have not been provided.
Google DeepMind demonstrated Genie 4 on October 27, showing an AI-generated scene that users can move through and change with text prompts. The company says the model creates real-time interactive worlds with sound, but it is not yet generally available, and DeepMind says objects and details can drift during sessions lasting only a few minutes.
DeepMind describes Genie 4 as a world model: rather than rendering a scene from a stored 3D map or producing a fixed video, it predicts what comes next from the frames already generated and the user’s inputs. In its demonstrations, users steer scenes with keyboard and mouse controls, then type instructions to alter elements such as weather, camera movement, or characters. The company says Gemini 3.8 Flash helps interpret goals, break them into steps, and monitor the scene.
The company also says Genie 4 generates audio alongside visuals, including ambient sounds, music, and dialogue, and simulates events such as collisions and aerodynamics. Those are vendor descriptions of capabilities shown in selected demonstrations; DeepMind has not published benchmark figures in the supplied announcement material. A robotics demonstration showed a simulated Apptronik Apollo humanoid playing checkers. It did not involve a physical robot, according to the source account.
Access is limited to trusted testers for now. DeepMind says the system is expected to come to Google AI Ultra subscribers through Project Genie and to early Gemini API partners, but has not supplied a launch date or API price. The source account reports that TechCrunch and Bloomberg, as cited by heise, described clips lasting up to about two minutes; that figure is not stated in DeepMind’s own text.
Infographic 01 · How Genie grew up
From flat sketches to worlds with sound, in under three years.
Each step as Google DeepMind described it. Session limits and consistency are different measures, so read the labels.
- Feb 2024
Genie 1Simple 2D worlds11 billion parameters, learned from unlabelled internet videos - Dec 2024
Genie 23D worlds from a single imageConsistent “for up to a minute”; real time only as a slimmed-down, lower-quality version - Aug 2025
Genie 3First real-time Genie720p at 24 frames per second, consistent “for a few minutes”, memory about one minute back - Jan 2026
Project GenieUS AI Ultra subscribers can play60 seconds per world, partly because of compute costs - May 2026Project Genie goes globalIn the $200 AI Ultra tier, for users 18 and older
- 27 Oct 2026
Genie 4Sound, live steering, more realismResearch preview for trusted testers; Ultra subscribers and API partners “coming soon”
Genie 3 vs. Genie 4
As described by Google DeepMind. Genie 4 column = vendor claims unless marked.
| Feature | Genie 3 (Aug 2025) | Genie 4 (Oct 2026) |
|---|---|---|
| Sound | Genie 3: None | Genie 4: Generated with the picture: ambient noise, music, dialogue |
| Steering | Genie 3: Keys, plus “promptable world events” | Genie 4: Free-form text changes while it runs; Gemini adapts and checks them |
| Look | Genie 3: 720p, 24 fps | Genie 4: “Unprecedented realism”; no resolution or frame rate published |
| How long | Genie 3: “A few minutes”; Project Genie capped at 60 s | Genie 4: “A few minutes” (DeepMind); up to ~2 min consistent scene reported |
| Access | Genie 3: Small preview, then Project Genie | Genie 4: Trusted testers; Ultra and API partners “coming soon” |
Sources: Genie paper, arXiv, 23 Feb 2024; Genie 2 blog, 4 Dec 2024; Genie 3 blog, 5 Aug 2025; TechCrunch, 29 Jan 2026; Google, 19 May 2026; Google DeepMind, Genie 4 blog, 27 Oct 2026; Genie 4 model page (limitations, disclaimer); TechCrunch, 27 Oct 2026. Table wording: Thorsten Meyer.
Powered by Thorsten Meyer AI
Where Interactive Worlds Could Matter
If Genie 4 can maintain responsive scenes for longer periods, it could give game developers and filmmakers a way to test ideas before building finished environments. DeepMind says tester feedback includes uses in game prototyping and film previsualization. Those are reported early applications, not proof of commercial results or a general release.
The technology also matters to robotics research, where generated environments may let teams simulate tasks. But the checkers demonstration was a simulation rather than a real-robot test, so it does not establish that Genie 4 can train robots effectively in practical settings. The system’s limited duration and scene drift also bear directly on whether it can support work that depends on stable spaces and repeatable results.
More realistic synthetic video and audio could make it harder for viewers to identify generated material. DeepMind says Genie 4 outputs carry an invisible SynthID watermark. That is a company statement about the system’s output; the source material does not provide independent testing of how reliably the watermark can be detected or retained.
Popular right now

Vitamix VX1 Professional-Grade Blender, 64 oz, Black

Bose QuietComfort Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones, Black

Samsung 55-Inch S90H OLED 4K UHD Smart TV Bundle with Full-Motion Mount and Surge Protector
As an affiliate, we earn on qualifying purchases.
From Short Clips to Live Worlds
Genie 4 follows a series of DeepMind systems with increasing interactivity. The first Genie, introduced in February 2024, generated simple 2D worlds. Genie 2, announced in December 2024, made 3D environments from a single image and maintained consistency for up to a minute, though real-time operation required a reduced-quality version.
In August 2025, Genie 3 was described as DeepMind’s first real-time world model, running at 720p and 24 frames per second and maintaining a scene for a few minutes. Google later made Genie 3 available to paying US subscribers through Project Genie, according to the source material. The progression gives Genie 4’s claims a practical frame: the aim is not just to generate a plausible clip, but to let users act inside a scene as it is produced.
Unlike a conventional game engine, which draws a view from a defined 3D environment, Genie 4 is described as predicting frames from prior visual information and user actions. That approach can produce flexible scenes without a conventional stored map, but it also helps explain the stated weakness: visual consistency may deteriorate over time.
“Details and objects can shift or drift over time.”
— Google DeepMind
Duration, Quality and Public Access
Several points remain unresolved. DeepMind has not published benchmark results in the supplied material, so the claims about realism, physics, and consistency cannot yet be compared against independent measurements. The company’s demonstrations were selected by DeepMind, and the source account notes that outside reports put some clips at about two minutes, while DeepMind’s own wording says only “a few minutes.”
It is also unclear how often scenes drift, how the model performs across different prompts, or whether users can reliably reproduce the same result. The stated upcoming access for Google AI Ultra subscribers and Gemini API partners has no announced date or API price. The source material does not establish how broad the trusted-tester group is or what restrictions will apply.
Potential uses in film, games, and robotics remain at different stages. Previsualization and prototyping are described as tester feedback, while the humanoid example is a research simulation. Neither establishes the model’s commercial readiness or real-world robotics value.
Awaiting Tester Access and Evidence
The next public milestones are likely to be more details on access through Project Genie and the Gemini API, along with information about pricing, availability, and usage limits. Google has not given a timetable in the supplied announcement account, so readers should treat “coming soon” as an undated plan rather than a scheduled launch.
As testers and outside observers gain access, the main questions will be whether Genie 4 can sustain coherent scenes beyond brief demonstrations, how well its sound and visual elements stay synchronized, and whether its outputs are useful for production or research tasks. Independent testing could also clarify the performance claims and the practical limits of SynthID marking. Until then, the clearest confirmed picture is a promising demonstration with short sessions, acknowledged drift, and restricted access.
Key Questions
What is Google Genie 4?
Genie 4 is a world model that DeepMind says generates scenes frame by frame in response to user actions and text instructions, rather than rendering a fixed video or using a conventional stored 3D map.
Can the public use Genie 4 now?
No broad public access has been announced. DeepMind says trusted testers can use it and describes future access for Google AI Ultra subscribers through Project Genie and early Gemini API partners, without giving a date.
How long can a Genie 4 world last?
DeepMind describes a few minutes of continuous interaction and warns that details or objects may drift. TechCrunch and Bloomberg, as cited by heise, reportedly described clips of about two minutes, but that number is not in DeepMind’s own wording.
Does Genie 4 train real robots?
The cited humanoid demonstration was a simulation of an Apptronik Apollo robot playing checkers, not a test with a physical robot. The source material does not show that Genie 4 has been validated for real-world robot training.
What has Google said about identifying Genie 4 outputs?
DeepMind says every Genie 4 video and audio output carries an invisible SynthID watermark. The supplied material does not include independent evidence about how reliably that mark can be detected.
Source: ThorstenMeyerAI.com
