Introduction

Google DeepMind has announced Genie 4, a real-time interactive world model that generates explorable environments frame by frame, now with natively generated audio and characters that can speak to you and to each other.

at 16:10 UTC on October 27 alongside the Genie 4 model page. You can't run it today. It's a research preview for trusted testers, with a waitlist to try it in Project Genie. Google says developer access through the Gemini API and AI Studio, and consumer access through Google AI Ultra, are "coming soon."

For game makers, this is the clearest look yet at where prompt-to-playable is heading. You type a world into existence, walk around in it, and keep rewriting it as you go: the weather, the camera, a new character. Gemini plans the next beat in the background.

What shipped

  • Model: Genie 4
  • Vendor: Google DeepMind
  • Modality: text (and reference image) to a real-time interactive world, with native audio and speech
  • Open vs API: closed. No weights. A Gemini API and AI Studio release is planned, with no pricing yet.
  • Where to run it today: nowhere public. Trusted testers only, plus the Project Genie waitlist. Phased rollout to AI Ultra is "coming soon."
  • Built on: Gemini Omni anchors appearance and behavior. A Gemini reasoning model (Ars Technica and The Verge both report Gemini 3.8 Flash) breaks prompts into plans and checks each output frame to keep the simulation coherent.

What's new over Genie 3

  • Native audio. This is DeepMind's first world model with sound. Footsteps crunch, crowds cheer, and you can write audio cues into the prompt. DeepMind's own examples end with lines like Audio description: Submerged bubbling sounds, gentle ambient water currents...
  • Speaking characters. Generated people move and react, and they can hold an adaptive conversation with you or with each other. For NPC work, this is the headline.
  • Live steering. Promptable events are now the main way you use it. Change the weather, move the camera or add elements in natural language mid-run, and Gemini rewrites the plan.
  • Physics and people. DeepMind claims better collisions, water flow and aerodynamics, and less uncanny human motion.
  • Longer sessions. About two minutes of continuous interaction, per The Verge, up from about one minute in Genie 3 and Project Genie.

Prompt shape to steal

DeepMind's showcase prompts read like scene docs, not one-liners. Each one has an Environment block, a Character block, camera rules and audio, and each one locks style hard. Here's a template that follows their pattern, ready for whenever you get access:

Environment: A rain-soaked night market on a hillside town, paper lanterns, steam from food stalls, wet cobblestones reflecting neon.
Character: First-person POV of a courier on foot, hands visible holding a parcel.
Camera: Smooth handheld walk forward; can turn freely into side alleys while staying coherent.
Style: Painterly stylized illustration at all times, no photoreal textures. No readable text anywhere.
Audio: Rain on canvas awnings, sizzling woks, distant chatter, vendors calling out.

Then steer it live. For example: a stall owner waves and asks where the parcel is going or the rain stops and fog rolls in.

Limitations DeepMind lists

  • Visual details and object permanence drift over longer runs and after many prompt edits.
  • Multi-agent interactions are still an open research problem.
  • It can't faithfully recreate real-world locations.
  • Text is only legible when you put it in the world description.
  • Sessions last a few minutes, and audio accuracy is still a work in progress.
  • All video and audio carries a SynthID watermark.

Why it matters for game devs

Genie 4 still isn't an engine. You don't get meshes, saveable state or exportable assets, so treat it as a pre-viz and prototyping surface, which is how Google pitches it alongside film pre-vis and robot training. What changes is the feedback loop: a designer can now hear a level and talk to a placeholder NPC before anyone builds it. The open-weight side of the lane has been racing on WASD control and saveable worlds (see WanGP's Hy-Nomas and EVOKE world generator and Tencent Hy-Nomas). Genie 4 pushes the closed frontier toward sound and dialogue.

We'll update when the Gemini API model ID and pricing land, or when the Project Genie waitlist opens up.

—Titus