DeepMind opened Genie 4 to eligible Google AI Ultra subscribers with live text edits and adaptive audio.
You do not regenerate to change a scene.
You type while you explore Project Genie to shift weather, add a character, or redirect action as visuals and sound rebuild around you, DeepMind says.
Google AI Pro members must join the waitlist on the same page.
Genie 3 stays in the model picker for longer runs, now up to 3 minutes per generation from 60 seconds with linked worlds, DeepMind says.
It also seeds worlds from 360-degree panoramas and fires real-time promptable events, shown as a flight over forests, underwater ruins, and icy seas that summons a rideable dragon on command.
The new Project Genie agent turns conversation into a built world with initial visuals and a custom background score from Google multimodal models.
Gemini can join the session to watch the screen, surface suggestion chips, or trigger events automatically as you travel.
The gallery ships with curated starters like Living Memory, Dino Field Guide, and Rock Paper Scissors where Gemini acts as referee, plus community builds you can remix.
Early creator builds include Villa Giallo with Will aka BLVCKL!GHT as a room-by-room gothic mystery and The Life of Ants with Anand Gandhi from an ant perspective, DeepMind says.
All video and audio carry SynthID watermarking and pass input and output safety screening, DeepMind says.
DeepMind lists remaining limits as camera latency, intricate physical interaction, real-time audio precision, and long-term scene permanence.
The same Genie 4 model also takes step text captions during rendering to vary robotics simulations for endless training, per model documentation.
If you ship client demos, plan for three-minute takes, remixable gallery rights, and no persistent physics yet.





