Google opens Genie 4 world model to AI Ultra subscribers, now with sound
Google on November 18 made its Genie 4 world model available in Project Genie to eligible Google AI Ultra subscribers, adding real-time audio, mid-session text edits, a world-building agent and a public gallery.
Cortex, covering AI models, labs and agents
, 2 min read
3 minGenie 3 generation length in Project Genie, up from 60 sec
What to know
- Genie 4 is available to eligible Google AI Ultra subscribers at labs.google/projectgenie; AI Pro subscribers can join a waitlist.
- Google says Genie 4 is the first Genie model to generate real-time native audio, and lets users change a running world with text prompts.
- Google published no resolution, frame rate, session length or benchmark figures for Genie 4, and there is no developer API yet.
Genie worlds used to be silent. The version Google put in front of paying customers on November 18 makes the rain sound like rain and lets characters talk, and you can tell it mid-session to change the weather or move the camera. That is the real change in Genie 4, a world model that generates a playable scene frame by frame as you move through it.
Google's post says Genie 4 is now in Project Genie, its Labs site for the model, for "eligible" Google AI Ultra subscribers. AI Pro subscribers get a waitlist. Google lists no countries or age limits; Unite.AI notes the first Project Genie in January was U.S.-only and 18-plus, and nothing says that has loosened. Genie 4 itself was shown as a research preview on October 27, so this is the first time people outside Google's testers can use it.
What you get
Google says Genie 4 is the first Genie model to pair its visuals with real-time native audio: ambient sound, music and character dialogue. Text prompts can edit a running world, swapping characters, changing the weather or redirecting the action. A Project Genie agent builds a world from a plain-language conversation and writes it a background score, and Gemini watches the session, suggests what to try and can trigger events. A community gallery lets people publish and remix worlds.
The older Genie 3 stays in the model picker with upgrades of its own: generations now run up to three minutes, up from 60 seconds, and worlds can be chained together.
What Google did not say
For a model launch, the post is thin on numbers. There is no resolution, frame rate, latency or maximum session length for Genie 4, and no benchmark of any kind. Google's own list of open problems is candid: camera latency, complex physical interactions, audio timing and keeping a scene consistent over time. Unite.AI and AlphaSignal both say detail and object permanence drift in long sessions or after edits. There is no API; AlphaSignal says early partner access through the Gemini API is planned, without a date.
So this is a toy for the most expensive tier, not a tool developers can build on. That matters because world models are where the labs are spending: AMD bought World Labs for its world-model work this month, and Google's pitch for this line since its August 2025 research model has been training robots and agents inside generated worlds. Letting Ultra subscribers poke at Genie 4 is a cheap way for Google to find where its worlds break. It arrived the same day as Gemini 4 Neon, which got the benchmarks and the price sheet this launch lacks.
Every frame and sound carries Google's SynthID watermark, and inputs and outputs pass through safety screening, Google says. What it has not said is when anyone without an Ultra plan gets in.
Sources
- Project Genie update featuring Genie 4, Google, Nov. 18, 2026
- Google Brings Genie 4 to Project Genie for Eligible AI Ultra Users, Unite.AI, Nov. 18, 2026
- Google DeepMind Opens Genie 4 to Ultra Subscribers With Real-Time Audio Worlds, AlphaSignal, Nov. 18, 2026
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.