Genie 4: Google DeepMind's new world model adds sound, live prompt edits and two-minute worlds
Quotable summary
Genie 4 is Google DeepMind’s newest world model: an AI that generates an interactive, explorable scene and simulates what happens next as you act in it. Announced on 27 October 2026, it adds native audio (music, ambient sound and character dialogue generated with the world), real-time edits by text prompt (change the weather, move the camera, add a person), and better physics and motion. TechCrunch and The Verge report scenes stay consistent for about two minutes, double Genie 3. DeepMind frames it plainly: “Understanding and simulating the world is a crucial aspect of building AGI.” For now it’s a research preview for trusted testers; Google AI Ultra subscribers and Gemini API early-access partners are next. Translation: AI labs are building practice worlds for robots and agents, and this one now talks back.
What changed
| Item | Genie 4 detail |
|---|---|
| Announced | 27 October 2026 (Google DeepMind blog, Google Keyword) |
| What it is | World model for real-time, interactive simulation |
| New vs Genie 3 | Higher photorealism; native audio; free-form prompt edits mid-simulation; better physics and more humanlike characters |
| Session length | About 2 minutes of consistent scene, up from about 1 minute (TechCrunch, The Verge) |
| Gemini’s role | DeepMind: Gemini Omni anchors appearance and behavior; Gemini 3.8 Flash acts as a guiding agent that breaks goals into steps, checks frames and steers the scene between your prompts |
| Access today | Research preview, trusted testers only |
| Next (“coming soon”; TechCrunch says weeks) | Google AI Ultra subscribers in Project Genie; early-access partners in the Gemini API and Google AI Studio; waitlist open |
Limitations DeepMind lists
| Limitation | What DeepMind says |
|---|---|
| Drift | Visual details and object permanence can shift as sessions run longer or after many edits |
| Multi-agent interaction | Modeling several independent agents in one world is still a research problem |
| Real places | Can’t simulate real-world locations with perfect accuracy |
| Text | Legible text usually appears only if it’s in the input description |
| Duration | A few minutes of continuous interaction, not long sessions |
Why it matters
- World models are AGI infrastructure. DeepMind’s stated reason for Genie is training robots and agents in safe, endlessly varied simulations. TechCrunch reports a Gemini Robotics experiment where a Gemini model broke a task into steps and Genie 4 simulated them live, a preview of testing robot plans before they touch the real world.
- Audio closes a big realism gap. Worlds that generate their own sound and dialogue are far more useful for training agents that must listen as well as see, and for filmmakers doing pre-visualization.
- Two minutes is still short. DeepMind is upfront that long sessions drift. Anything that needs long-horizon consistency, like a full game level or an hour-long robot rehearsal, isn’t here yet.
- Gated rollout. No public access yet, so all demo quality claims come from Google. Expect independent tests once Ultra subscribers get in.
FAQ
Can I try Genie 4 now? Only if you’re one of Google’s trusted testers. Everyone else can join the Project Genie waitlist. Google says Ultra subscribers and select Gemini API and AI Studio partners are next, with no firm date.
What’s a world model? An AI system that generates an environment and predicts how it changes in response to actions, so a person or an AI agent can move around and interact inside it in real time.
Which Gemini models are involved? DeepMind says Gemini Omni anchors appearance and behavior, and Gemini 3.8 Flash runs in the loop as a guiding agent that plans steps and checks each frame. In a separate Gemini Robotics experiment, TechCrunch reports a Gemini 4 Argon model planned the tasks Genie 4 simulated.
How is it different from Genie 3? More photorealistic, adds generated audio, accepts free-form edits mid-simulation, and holds a scene roughly twice as long.
Sources
| Source | Role | Link |
|---|---|---|
| Google DeepMind | Primary: announcement, capabilities, limitations, rollout | Google DeepMind — Genie 4: a new generation of world simulation |
| Primary: Keyword announcement | Google — Introducing Genie 4 | |
| Google DeepMind | Primary: model page and limitations | Google DeepMind — Genie 4 model page |
| TechCrunch | Secondary: briefing details, session length, robotics experiment | TechCrunch — Genie 4 adds audio and real-time steering |
| The Verge | Secondary: session length, rollout | The Verge — Google’s next Genie model |