🔍 Read the full analysis: Genie 4’S AI Worlds: What Works, What Breaks, And Who Has Access on ThorstenMeyerAI.com
Get smart everyday buys delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Google DeepMind presented Genie 4 on October 27, describing a real-time AI model that generates interactive visual worlds and audio. The model is not yet generally available: access is limited to trusted testers, with a future rollout to some Google AI Ultra subscribers and Gemini API partners announced without a date. DeepMind also acknowledges that scenes can drift and support only a few minutes of continuous interaction.
Google DeepMind introduced Genie 4 on October 27, describing an AI model that generates interactive scenes as users move through them and issue text instructions. The announcement matters because the system combines real-time visuals and audio, but access remains restricted to trusted testers, and DeepMind says worlds can drift during sessions lasting only a few minutes.
Unlike a video generator that renders a set clip, Genie 4 is presented as a world model: it predicts what appears next from the scene so far and from user inputs such as keyboard commands or typed requests. DeepMind says users can alter a running scene with instructions about weather, camera movement, or adding a person or object. The model does not use a stored 3D map or conventional game engine, according to the source material; it generates subsequent frames based on what has already been drawn.
DeepMind says Genie 4 generates sound alongside imagery, including environmental audio, music and character speech, and simulates interactions such as collisions and aerodynamic effects. It also describes Gemini 3.8 Flash as a guiding agent that can break tasks into steps and monitor the generated scene. Those capabilities are company descriptions, shown in selected demonstrations; the announcement material provided no independent benchmark results.
For now, DeepMind says Genie 4 is available to trusted testers. The company said access is expected to expand to Google AI Ultra subscribers through Project Genie and to early Gemini API partners, but did not provide a launch date or API pricing. The source reports Google AI Ultra at $200 per month at the time of writing. DeepMind says outputs carry an invisible SynthID watermark. The source also distinguishes claimed creative uses from research demonstrations: film previsualization and game prototyping are tester feedback, while the robot example was simulated rather than performed by a physical robot.
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
What Interactive Worlds Could Enable
If the described capabilities work reliably beyond curated demonstrations, Genie 4 could give developers and researchers a way to test ideas in changeable simulated environments without first building every scene as a conventional game. DeepMind has cited film previsualization and game prototypes as tester feedback, which could make the system relevant to creative workflows. Those reported uses are not evidence of broad commercial deployment or measured productivity gains.
The robotics example points to a separate research use: testing tasks in simulation before involving physical equipment. But the demonstration involved a simulated Apptronik Apollo humanoid playing checkers, not a real robot. The source says Waymo already uses the predecessor, Genie 3, for its driving simulator; that does not establish that Genie 4 is being used for Waymo operations.
There is also a public-facing concern as generated scenes become more realistic and include matching audio. DeepMind says it applies an invisible SynthID watermark to Genie 4 video and audio. The source does not describe how viewers can check for that watermark or how well it identifies outputs after they are shared or edited.
Popular right now
Bose QuietComfort Wireless Bluetooth Headphones with Active Noise Cancellation, Mic, USB‑C Charging, Deep Bass, 24‑Hour Battery – White Smok
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
Samsung 42-Inch Class OLED S90F 4K Smart TV (2025)
As an affiliate, we earn on qualifying purchases.
From Genie 3 to Real-Time Scenes
Google’s Genie models have developed from simpler generated environments toward interactive, real-time scenes. The source says the first Genie, introduced in February 2024, generated simple 2D worlds. Genie 2, announced in December 2024, could create 3D worlds from a single image and maintain consistency for up to a minute, though its real-time operation used a lower-quality version.
In August 2025, Genie 3 was presented as a real-time model running at 720p and 24 frames per second, with scenes remaining consistent for a few minutes and visual memory reaching back about one minute. Google opened Genie 3 through Project Genie to paying US subscribers in January 2026, according to the source. TechCrunch reported a 60-second limit per world, which it linked partly to computing costs.
The Genie 4 announcement builds on that progression with claims of more realistic imagery, native audio and live text-driven changes. DeepMind’s stated limits remain material: objects and details can shift or drift over time, and interaction lasts a few minutes. TechCrunch and Bloomberg, as cited by heise, reported scenes of up to about two minutes; the source notes that this specific duration does not appear in DeepMind’s own wording.
“Details and objects can shift or drift over time.”
— Google DeepMind
Limits, Testing and Access Still Open
There is not yet enough information in the source material to assess Genie 4 against independent tests. DeepMind has not published benchmark numbers, and the public examples were company-selected demonstrations. The source does not establish how performance varies across prompts, how reliably scenes preserve objects over time, or how often the model follows user instructions accurately.
Availability details are also incomplete. DeepMind has not supplied an opening date for Google AI Ultra subscribers or Gemini API partners, nor an API price. It is unclear whether the reported session length will be a fixed product limit, whether it will differ by access tier, or what compute requirements will apply. Reports of sessions lasting about two minutes are attributed to media coverage, not to the company’s announcement text.
The proposed uses need similar care. Previsualization and game prototyping are described as feedback from testers, while the humanoid example is a simulation. The source does not provide evidence that Genie 4 is ready for production robotics or that its generated environments are dependable substitutes for real-world testing.
Watch for Access and Independent Tests
The next practical milestone is Google’s planned expansion of access to AI Ultra subscribers through Project Genie and to early Gemini API partners. DeepMind has not announced dates, prices beyond the cited subscription tier, or detailed eligibility rules, so those terms remain to be confirmed.
Further information will also be needed to judge the model beyond its launch demonstrations: independent tests of visual consistency, interaction length, audio generation, instruction-following and watermark detection. Until those details and broader user experiences are available, Genie 4’s abilities should be treated as vendor claims and early demonstrations, not established performance across everyday use.
Key Questions
What is Genie 4?
Genie 4 is Google DeepMind’s described world model, which generates the next part of a visual scene as a user moves through it or enters instructions. DeepMind says it also produces audio alongside the imagery.
Can the public use Genie 4 now?
No general public release is described. DeepMind says access is currently limited to trusted testers, with planned availability for Google AI Ultra subscribers and early Gemini API partners. No date has been announced.
What are Genie 4’s known limitations?
DeepMind says details and objects can shift or drift over time and that the model supports a few minutes of continuous interaction. Media reports cited in the source describe scenes of up to about two minutes, but DeepMind’s announcement did not specify that exact limit.
Does Genie 4 train or control real robots?
The cited humanoid example was a simulation of an Apptronik Apollo robot playing checkers; no physical robot was involved. The source presents robot training as a research demonstration, not a confirmed production use of Genie 4.
How does Google identify Genie 4 media?
DeepMind says every Genie 4 video and audio output carries an invisible SynthID watermark. The source does not explain how people can inspect the watermark or how it behaves after media is edited or redistributed.
Source: ThorstenMeyerAI.com
