Overview
Genie 4 is the world model Google DeepMind announced on October 28, 2026, positioned as a real-time interactive simulation. Visuals and sound come from the same generation pass: as the world responds on screen, ambient audio, music and character dialogue are produced alongside it, and the people inside react and speak as events unfold.
The model is built on the Gemini family. Gemini Omni anchors simulations in believable appearance and behaviour, other Gemini models supply the reasoning that keeps a scene coherent, and Gemini 3.8 Flash runs inside the loop as a guiding agent that watches the frame, breaks goals into steps and checks progress. That loop is what holds multi-step causality together.
Interaction works like a game. Standard keyboard controls move the character and camera, and natural language rewrites the scene live by changing weather, moving the camera or introducing a person or object, with Gemini adapting the prompt to the surroundings. Genie 4 is in research preview with trusted testers, heading next to Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and AI Studio.
Key Features
- Native audio alongside video: Every world carries its own soundtrack: ambient noise, music and interactive dialogue are generated by the model rather than added afterwards, and characters speak as events happen.
- Steer scenes in real time: Change the weather, move the camera, add a character or an object with a sentence. Gemini rewrites the prompt against the current scene, so there is no pause to rebuild.
- Gemini in the loop: Gemini Omni anchors appearance and behaviour while Gemini 3.8 Flash observes the frame, decomposes goals and verifies progress, which is what keeps multi-step events causally consistent.
- Keyboard control and free camera: Navigate with standard keyboard input and steer the camera independently of the subject. DeepMind's demos use 16mm film grain and low-contrast natural light to test lighting and spatial continuity.
- SynthID on every output: Generated video and audio carry an imperceptible SynthID watermark, so material can later be identified as Google model output.
- Built for training environments: DeepMind points at safe training grounds for robots and agents, plus interactive experiences in film, journalism, gaming and manufacturing. It has already paired Gemini 4 Argon with Genie 4 to break user tasks into real-time steps for Apptronik's Apollo humanoid.
Use Cases
- Robot policy training in an environment that can vary without limit, covering collisions, grasping and navigation.
- Previsualisation for film and advertising, testing camera moves and lighting without waiting on a full render.
- Game level prototypes, changing terrain and weather in a sentence and seeing the interaction immediately.
- Interactive narrative experiments where characters speak in response to events in the scene.
- Evaluation environments for agents, giving tool use and long-horizon tasks a controllable world.
- Educational and newsroom explainers that need an editable simulation of a physical process.
Pros
- Video and audio come from the same pass, removing a separate dubbing and alignment step.
- Scenes are steerable with language while running, including camera and weather.
- Gemini 3.8 Flash in the loop stabilises causality across multi-step events.
- Keyboard and stick control feel close to a game, so exploration is not locked to a fixed path.
- All generated output carries SynthID, keeping provenance traceable.
- DeepMind has named concrete uses in robotics, film, gaming and manufacturing.
Pricing
Genie 4 is in research preview with trusted testers and is not sold separately yet. General access starts with Google AI Ultra subscribers through Project Genie, alongside early access partners in the Gemini API and Google AI Studio, with a waitlist on the Project Genie page. DeepMind has not published per-second or per-call pricing, nor API rate limits, so billing will follow whatever the company announces at general availability.
Summary
Genie 4 moves the world model from watchable to playable. Earlier versions concentrated on photorealism; this one adds real-time steering, audio generated with the frame, and Gemini keeping causality intact. Only with all three does the system work as a training ground or a production tool rather than a demo reel. It is still a research preview, so the practical question is access and pricing, not capability.
Version History
- Genie 4 launches with native audio and real-time language control (2026-10-28): Google DeepMind announced Genie 4, a world model in research preview with trusted testers. It is natively multimodal, generating ambient sound, music and character dialogue in step with the visuals; Gemini Omni anchors appearance and behaviour while Gemini 3.8 Flash acts as a guiding agent for multi-step causality. Keyboard and stick control plus natural language edits for weather, camera and objects. Every output carries SynthID. Next stop is Google AI Ultra subscribers in Project Genie and early access partners in the Gemini API and AI Studio, with no API pricing published yet.