A world model is not a video clip
A video generator renders a fixed sequence. Genie 4 simulates a place you move through: you navigate with standard keyboard controls, and the model predicts what happens next as you act. DeepMind describes it as generating how the world looks, sounds and responds, pixel by pixel, with no mesh or 3D export.
What is new in Genie 4
- Native audio: music, ambient sound and interactive dialogue generated alongside the visuals in one pass
- Live editing: change the weather, move the camera, or introduce a character or object with a text prompt while the simulation runs
- More natural people: DeepMind says human movement is more natural and physics, including collisions and aerodynamics, is stronger
- Longer coherence: reported scene coherence of about two minutes, roughly double Genie 3, though DeepMind itself does not publish a fixed session length
Two Gemini models steer the simulation. Gemini Omni anchors how the world looks and behaves in the real world, while Gemini 3.8 Flash acts as a guiding agent in the loop, checking steps and adapting the plan between your inputs. All generated video and audio carries DeepMind's SynthID watermark.
Who can actually try it
Genie 4 launched as a research preview for trusted testers, including filmmakers, game developers and robotics teams. Wider access is planned through Project Genie for Google AI Ultra subscribers and through the Gemini API and Google AI Studio for early-access partners, but DeepMind has not announced a firm public date or pricing.
| Genie 4 feature | What it means for you | Limit |
|---|---|---|
| Live text editing | Rewrite weather, objects and camera mid-scene | Edits can introduce drift. |
| Native audio | Dialogue and ambience match the scene | Audio accuracy is still imperfect. |
| Keyboard navigation | Explore instead of watching | No mesh or 3D export. |
A misty harbour town at dawn. Keep the stone lighthouse on the left. Now change the weather to driving rain and add a small fishing boat approaching the harbour, with gulls calling overhead.
Frequently Asked Questions
Is Genie 4 a game engine?
No. It is a research world model that simulates an explorable scene in real time. It does not export a game level, a mesh, or a playable build, and its consistency limits mean it is not a production pipeline yet.
When can the public use Genie 4?
DeepMind has not given a date. The stated path is trusted testers first, then Google AI Ultra subscribers through Project Genie and early-access partners through the Gemini API and AI Studio.
How long does a Genie 4 world stay coherent?
Press reports put coherent scenes at about two minutes, double Genie 3. DeepMind's own wording is vaguer — continuous interaction lasts a few minutes — and drift grows with session length and edit complexity.
Does Genie 4 generate sound as well as video?
Yes, and that is its headline change. It is DeepMind's first real-time world model to generate audio natively, including ambient sound, music and characters who talk back, all watermarked with SynthID.
