Google previews Genie 4 with native audio for interactive worlds

• Genie 4 generates interactive worlds with native audio, responsive dialogue and live scene edits, according to DeepMind's feature list.
• Gemini supplies world-state tracking and multi-step planning to keep scenes coherent, as Google's explanation describes.
• Trusted-tester access has begun; Google says Google AI Ultra access through Project Genie and early-partner API access are coming next.

A floating-saxophonist example specifies 16mm film grain, independent camera movement and no background music beyond the instrument in the scene. The same announcement's robotics experiment feeds instructions into a simulated humanoid's world as text captions while it renders.

DeepMind calls Genie 4 its first real-time world model to generate audio natively. Its announcement describes three kinds of output generated alongside the world:

• Environmental sound, including wind and waves.
• Music.
• Character dialogue that responds to user inputs and in-world events.

Live edits and character dialogue in Genie 4

Gemini Omni anchors simulations in real-world appearance and behavior, according to Google's overview. The improvements claimed in DeepMind's rundown cover:

• Aerodynamics.
• Collisions.
• Natural human movement and reactions to the environment.

Keyboard controls handle navigation and camera movement; text prompts modify the running simulation, as DeepMind's demonstrations describe. Overly static worlds were a limitation of Genie 3, one developer said, highlighting prompted events as an improvement in the new version.

A tester generated a city inhabited by robots and horses, then added dragons and an otter NPC without leaving the session.

Gemini 3.8 Flash gets the director's chair, supervising generated scenes through a feedback loop detailed in DeepMind's technical description:

1. Observe the scene as Genie generates it.
2. Break the requested goal into steps.
3. Review incoming frames to track progress and check whether each step succeeded.
4. Revise the plan when the world changes or the user intervenes.

That frame-based check is the standout engineering detail.

Google also paired Genie 4 with Gemini 4 Argon in research bringing together its world-model work and Gemini Robotics. In the Apollo demonstration, Argon breaks a user's task into steps for Genie to render, producing a simulation of Apptronik's humanoid playing checkers.

Sessions last up to two minutes, according to one launch summary. DeepMind's limitations section identifies several remaining problems:

• Persistence: Visual details and object permanence can drift during longer runs, navigation or repeated complex prompt edits.
• Physics: Complex physical interactions remain an active research area.
• Audio: Real-time audio accuracy still needs work.
• Rendering: Legible text and spatial accuracy remain difficult.

One hands-on tester described the experience as “glitchy” while demonstrating a mouse growing wings and flying to a teacup.

Google is gathering feedback through a research preview before expanding access. Its rollout announcement names these next destinations:

• Google AI Ultra subscribers through Project Genie.
• Early-access partners through the Gemini API.
• Early-access partners through Google AI Studio.

DeepMind says all generated video and audio will carry SynthID watermarks. It also reports internal safety evaluations before the preview and safety guardrails aligned with its policies.