Google Genie 4: The World Model That Talks Back — and Rewrites Itself Mid-Run

How it unfolded

Genie 4 is the fourth generation of DeepMind’s world-model line. Its predecessor, Genie 3, entered research preview in 2025 with real-time navigation and event triggers and scenes that held together for about a minute; Genie 4 pushes that figure to about two minutes and adds the native audio its predecessor lacked. In January 2026, Project Genie turned this kind of model into a web prototype for a wider audience, and Genie 4’s wider release will also arrive through Project Genie.

This update consists of four capabilities. A running world can be rewritten in natural language: weather, camera angle, a new person or a new object can all join while the simulation continues, with Gemini adapting each prompt to the surrounding scene and steering the scene between inputs to keep it coherent. On physics and people, the model simulates physical processes such as aerodynamics and collisions, characters move more naturally and can be steered in real time, and DeepMind calls the people simulation the largest leap of this release. On audio, Genie 4 is DeepMind’s first real-time world model to generate audio natively: wind, waves and character dialogue change live with inputs and events, rather than being dubbed on afterwards.

The steering loop is run by Gemini 3.8 Flash: inside the loop it observes the scene, breaks a goal into steps and checks progress frame by frame, adjusting its plan as the scene changes, while Gemini Omni anchors appearance and behavior in the simulation. DeepMind demonstrated Gemini 4 Argon combined with Genie 4 simulating Apptronik’s Apollo humanoid playing a game of checkers: Argon breaks the task into steps, the steps are interleaved as text captions into Genie 4, and the simulation adjusts as the task advances.

Rollout comes in three steps: the research preview goes to trusted testers first, then to Google AI Ultra subscribers in Project Genie and early access partners for the Gemini API and Google AI Studio, with a Project Genie waitlist already open. DeepMind says the trusted testers come from five fields: creative work, journalism, education, gaming and engineering.

Statements

DeepMind blog authors Raia Hadsell, Diego Rivas and Hang Qi write that understanding and simulating the world is a crucial aspect of building AGI; Genie 4 aims to produce safe, infinitely variable simulations for robot and agent training environments as well as new interactive experiences.

Tester use cases come from DeepMind’s tester summary: creative teams treat Genie 4 as a pre-visualization draft mode, shooting several takes inside a simulation; historical-scenario teams test turning a single 2D archival photograph into an environment simulation; game teams test the seconds-long jump from an idea to an explorable prototype; enterprise teams evaluate rapid scenario generation and point out gaps to close.

DeepMind lists the limits in the same blog post: Genie 4 remains an early experimental model, and after long runs, several complex edits or travel across a scene, visual details and object permanence can drift, while complex physical interaction, real-time audio accuracy, legible text rendering and spatial accuracy remain under research. The next direction is hybrid world models: Gemini plans structure and narrative timeline in the background, while Genie renders the immediate physical reality pixel by pixel.

Independent coverage matches the official blog on two key figures: Silicon Report and Skimwire both report scenes holding together for about two minutes, double Genie 3, and a rollout that starts with the research preview before reaching Ultra subscribers and API early access partners. Both reports list native audio and live rewriting as the main changes of this release.

Key facts

1. Launch: Google DeepMind announced Genie 4 on Oct 27 in its official blog, authored by Raia Hadsell, Diego Rivas and Hang Qi; the research preview opens to trusted testers first.
2. Form: a real-time interactive simulated world, with music, ambient sound and interactive dialogue generated in sync with the picture, not a video clip dubbed afterwards.
3. Live rewriting: a running world can be rewritten in natural language — weather, camera, objects and people — with Gemini adapting prompts to the scene and steering continuously.
4. Steering loop: Gemini 3.8 Flash observes the scene, breaks down goals and checks progress frame by frame inside the loop; Gemini Omni anchors appearance and behavior; the demo includes Gemini 4 Argon with Genie 4 simulating the Apollo humanoid playing checkers.
5. Scene length: scenes hold together for about two minutes, double Genie 3’s about one minute; the figure comes from launch-day independent coverage, while the official blog says coherence holds over short sessions.
6. Rollout: after the research preview, Google AI Ultra subscribers in Project Genie and early access partners for the Gemini API and Google AI Studio follow; the waitlist is open.
7. Watermark and limits: all generated video and audio carry a SynthID watermark; details and object permanence can drift after long runs, and complex physical interaction, real-time audio accuracy, legible text rendering and spatial accuracy remain under research.