Google Genie 4 world model adds native audio and real-time steering: preview only
DeepMind's Genie 4 generates steerable interactive worlds with built-in sound; it is a trusted-tester preview with no price or date for wider access.
· 5 min read
Written by AI from the sources linked below. How we use AI
On this page
Google DeepMind announced Genie 4 on 27 October 2026, a world model that generates interactive, steerable simulations in real time and, for the first time in the Genie line, produces audio natively. It is a research preview for a limited group of trusted testers; Google says it is coming soon to Google AI Ultra subscribers in Project Genie and to early-access partners in the Gemini API and Google AI Studio. No price and no date were given.
Key points
- Genie 4 was announced on 27 October 2026 as a research preview for trusted testers, with a waitlist open.
- Google calls it its "first real-time world model to generate audio natively": ambient sound, music and interactive dialogue.
- Users steer a running world with keyboard controls and natural-language prompts, such as changing the weather, moving the camera or adding characters.
- TechCrunch reports scenes stay consistent for up to two minutes, double Genie 3's limit; Google's own post says only that coherence holds "over short sessions".
- All generated video and audio will carry SynthID watermarks, per Google.
What happened
In its announcement, Google describes Genie 4 as a model for "real-time interactive simulation". It builds on the Gemini family: Gemini Omni anchors each world in real-world appearance and behaviour, while other Gemini models supply the reasoning needed to keep a simulation coherent and continuous. You set an event in motion and then steer the scene as it runs.
The longer DeepMind post adds that steering combines standard keyboard controls with text prompts, with Gemini adapting those prompts to the scene and keeping the world moving between inputs. In an agentic setup, it says, Gemini 3.8 Flash observes frames, breaks a goal into steps, checks progress and replans when the world changes or the user intervenes.
TechCrunch describes the changes from Genie 3 as audio that matches the scene, more photorealistic worlds with better physics and more human-like characters, and the ability to type prompts mid-simulation. It also reports a Gemini Robotics experiment in which a model it calls "Gemini 4 Argon" split tasks into steps that Genie 4 simulated in real time, as a way to test robot scenarios before real-world deployment. Google's own post names Gemini 3.8 Flash for the agentic role instead; the two accounts may describe different parts of the demo, and we could not reconcile them.
Background
Genie is DeepMind's line of generative world models. Genie 3 launched in August 2025 as a research preview, with output at 720p and 24 frames per second according to Wikipedia's summary, which we treat as secondary. On 29 January 2026 Google opened Project Genie to Google AI Ultra subscribers in the US as an "experimental research prototype". TechCrunch's hands-on at the time found a 60-second limit on world generation and navigation because of compute constraints, since Genie 3 is auto-regressive and needs dedicated compute for each session. It also found inconsistent results, characters that sometimes walked through walls, unresponsive keyboard controls and imperfect memory: a revisited area produced a second mug once. Guardrails blocked nudity and copyrighted characters.
Those limits are the yardstick for Genie 4. A world model that can only hold together for a minute is a demo; one that can run longer, respond to text mid-run and make sound is closer to a tool.
How it works
Neither Google post publishes a technical report, parameter count, resolution, frame rate or latency, and we have not seen any. What they describe is a division of labour. The Genie model generates frames and audio; Gemini models sit around it. Gemini Omni grounds appearance and behaviour in the real world, and Gemini reasoning models handle planning and coherence. For an agent, that loop is: observe the generated frames, decide the next step, act through the same controls a person would use, check whether the world changed as expected, and replan.
That loop is why Google pitches Genie 4 at training and testing agents. The announcement says simulating the world is "a crucial step toward AGI", enabling safe training environments for robots and agents, and it lists filmmaking, journalism, gaming and manufacturing as creative applications. TechCrunch adds interactive learning and says trusted testers see potential in film prototyping, particularly ideation and storyboarding.
What it means
For developers. There is nothing to call yet. Access goes first to trusted testers, then to AI Ultra subscribers in Project Genie, then to Gemini API and AI Studio early-access partners. Anyone planning around it should join the waitlist and wait for documentation, pricing and rate limits, none of which Google has published.
For game and film teams. Native audio makes a generated world feel more finished, but Google's own list of open problems is long: visual details and object permanence can drift, especially in longer sessions or after several edits, and complex physical interactions, real-time audio accuracy, legible text, spatial accuracy and long-term consistency are all listed as areas of active research. That points to previsualisation and prototyping rather than final assets.
For the competition. TechCrunch notes that Runway's GWM Worlds 2, released about two months earlier, is also playable, responds to user input and generates audio. It lists further deal activity in the field: AMD is acquiring World Labs, Amazon is reportedly in talks to buy Decart, and Rivian acquired Third Dimension for self-driving simulation. Interactive world models are becoming a category rather than a single lab's experiment.
For trust. Google says Genie 4 has safety guardrails in line with its policies across models, was internally evaluated before the preview, and that all generated video and audio will include SynthID watermarking. Detection of that watermark depends on Google's tools, so it helps with provenance only where those are used.
What we don't know yet
- Resolution, frame rate, latency and hardware requirements; neither Google post gives them.
- The real session limit. The "two minutes" figure is TechCrunch's, and Google's post only says coherence holds over short sessions.
- Pricing, and whether Project Genie will keep a per-session time cap.
- How the Gemini model names in the robotics demo fit together (Gemini 3.8 Flash in Google's post, "Gemini 4 Argon" in TechCrunch's account).
- Independent results: so far everything comes from Google's demos and trusted testers.
What to watch next
The Project Genie rollout to AI Ultra subscribers, which Google says is coming soon, will give the first outside hands-on tests of audio quality and drift. Watch for a technical report, Gemini API documentation, and whether robotics teams report that policies trained or tested in Genie 4 worlds transfer to real machines. Google has not claimed that they do.
Sources
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Google GTIG: exploit_agent turns patches into exploits, no human involved
Google's threat team says attackers ran an AI framework that diffs patches, writes exploits and runs them automatically, the first fully autonomous pipeline it has seen in the wild.
audiotale.net team8 min read
A2UI 1.0 released: Google's stable protocol for agent-generated native UIs
Google calls A2UI 1.0 the first stable version of its open protocol that lets AI agents describe screens and apps render them natively, with renderers for React, Angular, Flutter and SwiftUI.
audiotale.net team6 min read
Google Gemini glasses: pre-orders open from $395, on sale 10 December in US and Korea
Google and Samsung's first Gemini glasses, with Warby Parker and Gentle Monster frames, have camera, speakers and no display, and Google leads with privacy protections for bystanders.
audiotale.net team5 min read