Genie 4 reaches Ultra subscribers with live world editing
Edited and reviewed by AI Ping Editorial Team
3-line summary
- Project Genie is rolling out to eligible Ultra subscribers, while Pro members can join a waitlist.
- Genie 4 outputs synchronized audio and 720p video at up to 24fps, with around one minute of continuous interaction.
- Genie 3 gets up to three-minute sessions and panorama inputs; enabling remix exposes source assets.
Google has begun opening Genie 4 to eligible Google AI Ultra subscribers through an updated Project Genie experience announced on November 18. Users can change weather or characters while exploring a generated world, with audio adapting to the action. Navigation and new prompts help determine what appears next.
Access requires a personal Google Account, age 18+, and a supported Ultra region. Pro members can join a waitlist. The early research prototype consumes no AI credits for worlds, but daily limits depend on plan and demand; queues may apply.
The Genie 4 model card specifies synchronized audio and 720p video at up to 24 frames per second. Inputs include text, images, and live keyboard or mouse actions.
The current model overview lists around one minute of continuous interaction. Objects and visual details can drift during longer exploration or repeated complex edits. Complex interactions between users, multiple agents and their surroundings remain a research challenge. Accurate real-world locations, legible text and precise live audio remain limitations. Google labels its displayed recordings as real-time interactions, with sequences shortened for illustration.
Genie 3 remains in the model picker and receives a separate increase from 60 seconds to up to three minutes per generation. That update also supports chaining worlds and starting from 360-degree panoramas. These longer sessions belong to Genie 3; Genie 4's current overview describes the shorter duration.
Google's prompting guide describes an image preview before exploration begins. Users can revise its framing or supply their own starting image, and choose a first-person or third-person view. The guide covers painterly or claymation styles as well as realism, and uses foreground objects and specific surface sounds as examples of grounding a scene. For live steering, the guide recommends short requests for one change at a time. It also says dialogue works best with a single character, a constraint relevant to scenes involving conversation.
Sharing can expose the building materials
The Project Genie agent develops an idea through conversation while generating visuals and audio. The community gallery lets creators remix environments, characters and sound from existing experiences. Google's showcased examples include The Life of Ants, following a worker ant's activities, and Villa Giallo, a room-by-room mystery with clickable clues. Dream Flight changes the surrounding world when a user selects a world during flight.
Sharing with remix enabled exposes reusable prompts, images and source assets, according to Google. The gallery shows the submitted version and account name; updates require republication. Creators can make experiences private or disable remixing. The FAQ says Google will not claim ownership of original generated content, while its Terms of Service govern use.
The labs.google/fx FAQ also describes how interactions, outputs, usage information and feedback are collected. Google says automated tools help strip identifying details such as email addresses and phone numbers. Interactions sent to trained reviewers are separated from the Google Account and retained for up to 18 months. Switching history off excludes future interactions from human review and generative-model improvement, but activity can remain associated with the account for up to 72 hours to provide the service and process feedback.
The Genie 4 model card lists training-data filtering, product safety filters and SynthID watermarking among its safeguards. Google describes specialist red teaming outside the development team and safety reviews before release. Generated video and audio carry SynthID to help identify AI-created material.
Sources
- Genie 4 is now available — November 18 announcement — Google
- Get started with Project Genie — account eligibility, credits and sharing — Google Labs Help
- Genie 4 model card — published November 18, 2026 — Google DeepMind
- Genie 4 — current duration, limitations and recording disclaimer — Google DeepMind
- How to create effective prompts with Genie — Google DeepMind
- Project Genie — agent, gallery and showcased experiences — Google Labs
- labs.google/fx FAQ — review, retention and history controls — Google Labs