Genie 4 arrives in research preview with native audio, real physics, and Gemini-powered steering

Google DeepMind has released Genie 4 as a research preview for trusted testers, with broader access coming first to Google AI Ultra subscribers via Project Genie and early partners on the Gemini API and Google AI Studio. The release marks the fourth generation of DeepMind’s Generative Interactive Environments series and brings three capabilities that did not exist in any previous version: native audio generation, photorealistic physics simulation, and live scene steering powered by Gemini models.

What Genie actually does

Genie models generate interactive virtual worlds in real time from a text description or image prompt. Rather than producing a finished video, the model calculates the scene frame by frame and reacts to input as it goes. Think of it less like a film renderer and more like a physics engine that has watched everything ever recorded and can construct plausible environments on demand.

The first version launched in March 2024. Genie 2 followed in December 2024 with 3D environments. Genie 3, released in August 2025, pushed resolution and extended visual consistency to a few minutes. On 29 January 2026, DeepMind opened Project Genie to AI Ultra subscribers as the consumer access point for the series.

Genie 4 is a meaningful step beyond all of that.

The three new capabilities

Native audio. Every world Genie 4 generates carries ambient sound from the moment it spawns: wind, water, crowd noise, material acoustics. More significantly, Genie 4 introduces adaptive real-time character dialogue. Simulated characters converse and react to user inputs or to events elsewhere in the scene, without the interaction pausing to buffer or re-render. No prior real-time world model had anything equivalent.

Photorealistic physics. Genie 4 adds a credible simulation of real-world physics: aerodynamics, collisions, fluid behaviour. DeepMind’s demonstration footage showed a watermelon splitting with physically accurate fragment dispersion, a maple seed spinning downward in autorotation, and a dense Tokyo rail-platform crowd moving with individual-level behavioural variation, all generated live. The model also takes a leap forward in how it handles people: characters move naturally and respond organically to their environments rather than repeating looped animations.

Gemini-powered steering. You can set an event in motion and then guide it using natural language: shift the camera angle, change the weather, introduce a new character. For multi-step goals, Gemini 3.8 Flash operates as a planning agent inside the loop. While Genie 4 renders the world, Gemini continuously observes incoming frames, breaks goals into steps, verifies that each step succeeded, and adapts if something changes. This is what DeepMind means when it talks about “reliable causality” and “procedural behaviours”: the simulation does not just look plausible, it follows through on cause and effect.

Genie 4 can maintain a consistent scene for up to two minutes, double the limit of Genie 3.

How the Gemini models fit together

Several Gemini models are working inside Genie 4 simultaneously. Gemini Omni anchors environments in real-world appearance and behaviour. Gemini 3.8 Flash tracks state and plans next steps. In one demonstration, Gemini 4 Argon was paired with Genie 4 to simulate Apptronik’s Apollo humanoid robot playing checkers in real time, decomposing the task step by step and verifying each move.

That last point is the one to watch. DeepMind frames the robotics integration as an early experiment, but the underlying idea is direct: Genie 4 could serve as a synthetic training environment capable of generating physically accurate, behaviourally complex scenarios at a scale and variety that real-world data collection cannot match.

What this means for you

If you are a Google AI Ultra subscriber, access to Genie 4 through Project Genie is coming in the weeks following the research preview. If you are building on the Gemini API or AI Studio, early-access partnerships are also opening up.

The practical uses already being tested with trusted testers across creative industries, journalism, education, gaming, and engineering include:

• VFX pre-visualisation for film and TV production, letting teams rough out scenes before committing to expensive shoots
• Historical reconstruction, bringing archival moments to life inside a steerable environment rather than a static video
• Dynamic environmental modelling, including flash floods and wildfires, for education or emergency planning
• Game design prototyping, generating interactive environments from concept art or text briefs

For developers, the interesting angle is the Gemini API integration. Genie 4 runs on Google-hosted infrastructure through Project Genie and the Gemini API, so access does not require any local compute. The frame-by-frame rendering approach also means that downstream applications can intercept and act on the world state at any point, which opens up agentic use cases beyond just watching a simulation play out.

Known limitations

DeepMind is transparent that this is a research preview. Visual details and object permanence can still drift over time, particularly after multiple complex prompt edits or extended navigation through a scene. Legible text rendering, spatial accuracy, long-term consistency, and real-time audio precision remain active areas of ongoing research.

Google research director Tejas Kulkarni noted during a briefing that future hybrid world models could combine Gemini’s code and text generation with Genie’s real-time simulation at different levels of abstraction. That gives a reasonable sense of where the roadmap is heading, even if Genie 4 is not the finished destination.

The broader framing from DeepMind is that understanding and simulating the world is a core part of building AGI. The consumer-facing features are real and useful, but the robotics and agent training work is what explains why DeepMind keeps investing heavily in this line. Genie 4 is both a tool you might use and a building block for something considerably larger.