Genie 4 extends Genie 3 with audio, dialogue and more lifelike people, and TechCrunch reports scenes stay consistent for up to 2 minutes, double Genie 3's limit. It is a research preview for trusted testers, with Google AI Ultra via Project Genie next, so filmmakers, game designers and robotics researchers can join the waitlist but cannot build on it yet.
What changed
Genie 4 is Google DeepMind's real-time interactive world model, announced on 27 October 2026. It generates video and audio together, including ambient sound and spoken dialogue, and lets users steer a running scene with text prompts. DeepMind positions it for robot and agent training environments and for new interactive media.
Provider: Google DeepMind · Family: Genie
More about Google DeepMind on HokAI
Input modalities: text, image · Output: video, audio
About Genie 4
Genie 4 is the world model Google DeepMind announced on 27 October 2026, written up by Raia Hadsell, Diego Rivas and Hang Qi. It succeeds Genie 3, which DeepMind's blog lists as an August 2025 release, and it sits inside the Google DeepMind model line next to Gemini Omni, Veo 3.1 and the Lyria 3.5 music model, all grouped on the DeepMind provider page. DeepMind says it builds on Gemini capabilities: Gemini Omni anchors each simulation in real-world appearance and behavior, while other Gemini models supply the reasoning that keeps a scene coherent from one moment to the next.
The user steers a running world with keyboard controls for movement and camera, and with text prompts that change the weather, move the camera or add a person or object. Genie 4 also generates sound with the picture, from ambient noise and music to dialogue that responds to what the user or the world does. DeepMind describes it as its first real-time world model to produce audio natively. On physics, it names aerodynamics and collisions, and it says characters now move and react more naturally than before.
A second layer pairs Genie with Gemini as a director. In DeepMind's description, Gemini 3.8 Flash watches the incoming frames, splits a goal into steps and checks that each step happened. In a robotics experiment, Gemini 4 Argon broke a task into steps that Genie 4 rendered, with Apptronik's Apollo humanoid playing checkers. TechCrunch adds that a scene stays consistent for up to two minutes, double what Genie 3 managed, and notes that Runway shipped its own world model, GWM Worlds 2, a couple of months earlier (see Runway).
What is missing matters for anyone planning around it. DeepMind published no benchmark scores, resolution, frame rate, latency, parameter count or price, and the model is a research preview for trusted testers, so it sits with the other preview-stage models and with the proprietary models. Google says it is coming soon to Google AI Ultra subscribers in Project Genie and to early access partners of the Gemini API and Google AI Studio, with a waitlist open now. DeepMind's own limitations list is candid: details and object permanence can drift in longer runs, and complex physical interactions, real-time audio accuracy, legible text and spatial accuracy remain open research. Trusted testers in film, journalism, games and engineering are using it for previsualization drafts, explorable game concepts and scenario generation, which fits a draft tool better than a production one today.
Pricing
No public price. Genie 4 is a research preview for trusted testers, and Google has announced no API pricing as of 2026-10-29. The announced consumer route is Google AI Ultra through [Project Genie](/hub/tools/project-genie), whose own record tracks that plan.
Key Features
- Native audio and dialogue: Ambient sound, music and interactive speech are generated with the world, and characters respond to user input in real time.
- Text-steered live scenes: Prompts change weather, camera angle, people and objects mid-run, and Gemini adapts each prompt to fit the scene.
- Physics and people: DeepMind names aerodynamics and collisions as simulated, and characters can be steered while moving and reacting naturally.
- Gemini in the loop: Gemini 3.8 Flash breaks goals into steps and checks incoming frames to verify each step succeeded, keeping multi-step events on track.
- SynthID watermarking: All generated video and audio carries SynthID, DeepMind's imperceptible watermark for identifying AI content.
Pros
- Doubles the reported scene-consistency window of Genie 3 to up to two minutes, the clearest measured gain over its predecessor (TechCrunch).
- One model covers picture, sound and speech, so a draft scene does not need a separate audio tool.
- The Gemini director loop gives DeepMind a way to test multi-step task plans in simulation before real-world trials, as the robotics checkers experiment shows.
Cons
- No public access, API or price yet: trusted testers only.
- No published benchmarks, resolution, frame rate or latency, so it cannot be compared with rivals on numbers.
- DeepMind itself flags drift, unreliable text and imperfect audio as unsolved.
Frequently Asked Questions
How much does Genie 4 cost?
Nothing is priced yet. Google lists Genie 4 as a research preview for trusted testers and has published no API rate, per-second price or subscription fee for it. The Ultra subscription is the first planned entry point, so check the Project Genie record for current plan notes before budgeting.
How does Genie 4 compare to Genie 3?
Genie 4 adds native audio with spoken dialogue, text edits to a live scene, and more lifelike people and physics. TechCrunch reports scenes hold together for up to two minutes, double its predecessor. Pick Genie 4 for anything that needs sound or mid-scene changes, and note that DeepMind has not published side-by-side benchmark numbers.
Is Genie 4 open source, and can anyone use it yet?
It is proprietary and not generally available. Access is limited to trusted testers, with Google AI Ultra subscribers in Project Genie listed as next, followed by Gemini API and AI Studio partners. A waitlist for Project Genie is open on DeepMind's announcement page.
Does Genie 4 mark what it generates as AI?
Yes, DeepMind says all video and audio from Genie 4 carries SynthID, its imperceptible watermark. It also says the model went through internal safety evaluations before the preview and was released to trusted testers first. Details of the test results have not been published.
Who should wait for Genie 4, and who should look elsewhere?
Filmmakers drafting scenes, game designers sketching explorable concepts and robotics researchers testing scenarios are the groups DeepMind's testers describe. If you need a finished, non-interactive clip rather than a live world, Veo 3.1 and Gemini Omni are the existing Google options. Skip Genie 4 for work that needs readable on-screen text or long, stable sessions.
Top Alternatives
- Gemini Omni: Pick Genie 4 to explore and steer a live world with sound; pick Gemini Omni to generate or conversationally edit a finished clip.
- Veo 3.1: Pick Veo 3.1 when you need a finished video clip; pick Genie 4 only once you are accepted into the preview and want a world you can steer.