# Google DeepMind Genie 4 adds native audio to world model

URL: https://techaiwire.com/articles/google-deepmind-genie-4-world-model-native-audio-gemini-api/

_By Claire Whitmore - Oct 28, 2026_

Google DeepMind announced Genie 4 on October 27, 2026, a world model that generates interactive scenes with sound built in. A world model is an AI system that creates a place you can move through and change in real time, like a game level made on the fly. [Google's announcement](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4) says every world now "generates native audio on the fly, including ambient sounds, music, and interactive dialogue." A small group of trusted testers has it now, and developers are next in line through the Gemini API.

Google frames the work as more than entertainment. "Simulating the world is a crucial step toward AGI, enabling safe training environments for robots and agents," its blog post says.

## What Genie 4 adds

Users steer a scene with plain-language instructions while it runs. Google's examples include moving the camera, changing the weather and adding characters.

The [Google DeepMind model page](https://deepmind.google/models/genie/) calls Genie 4 "our most realistic and interactive world model yet." It says Gemini Omni anchors how things look and behave. Other Gemini models handle the reasoning that keeps a scene coherent over time.

[Gizmodo](https://gizmodo.com/googles-new-ai-world-model-can-generate-human-voices-not-just-video-2000831335) compares it with Genie 3, which Google unveiled in August last year. Genie 3 had no audio, and its people "tended to be more mechanical," the outlet writes. Genie 4 shows smoother human motion and facial expressions, with characters that talk and react. Gizmodo's headline says the model generates human voices. Google's own pages describe "interactive dialogue" but do not use the word voice.

## Who can use it, and when

| Group | Access |
|---|---|
| Trusted testers | Rolling out now as a research preview |
| Google AI Ultra subscribers | "Coming soon" in Project Genie |
| Gemini API early-access partners | "Coming soon" |
| Google AI Studio early-access partners | "Coming soon" |
| Everyone else | A waitlist on the DeepMind page |

Google has not given a price or a date for wider release. Gizmodo reports the current testers include video game developers.

Project Genie is Google's web app for exploring generated worlds. Gizmodo notes that its launch in January reportedly sent video game stocks sharply lower.

## Safety and watermarks

Google DeepMind says it released Genie 4 to trusted testers first on purpose. The model has "strict safety guardrails that are consistent with our safety policies across other models," the model page says.

Every video and audio clip Genie 4 makes will carry SynthID, Google's invisible watermark for spotting AI-generated content. That covers both picture and sound.

## What Google says it cannot do yet

The model page has a frank list of limits:

- It supports "a few minutes of continuous interaction, rather than extended sessions."
- It cannot simulate real-world locations with perfect accuracy.
- Real-time audio accuracy is still being improved.
- Visual details and object permanence can drift over time, especially after several complex prompt edits.
- Interactions between many independent agents in a shared world remain hard.
- Legible text usually appears only when the input describes it.

Gizmodo adds that the model "occasionally struggles to maintain visual consistency and audio accuracy."

## What this means for developers

Genie 4 is not something most developers can call today. The steps worth taking now are about preparation.

- **Join the waitlist if you build games or simulations.** Early-access Gemini API partners will shape how the model is exposed. Getting in early gives you a say.
- **Design for short sessions.** A few minutes of interaction suits prototypes, demos and level ideas. It does not yet suit long play sessions or persistent worlds.
- **Treat it as a robot training ground, carefully.** Google is exploring robotics uses. Worlds that drift over time can teach agents the wrong lessons, so check results against real-world tests.
- **Plan for watermark checks.** All output carries SynthID. If your product mixes real and generated footage, decide how you will label it and detect it.
- **Hold your budget.** There is no published price yet. Avoid building a business case until rates and rate limits are public.
- **Watch the audio.** Sound is the headline feature and also a listed weak spot. Test whether dialogue stays in sync before you rely on it for really interactive characters.

Genie 4 moves world models from silent clips toward scenes with sound and speech. The few-minute limit and the drift problems show how far they still are from replacing a game engine.

## Sources

- [We're announcing Genie 4, our new world model for real-time interactive simulation.](https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-genie-4) - Google
- [Genie 4](https://deepmind.google/models/genie/) - Google DeepMind
- [Google's New AI World Model Can Generate Human Voices, Not Just Video](https://gizmodo.com/googles-new-ai-world-model-can-generate-human-voices-not-just-video-2000831335) - Gizmodo