In this article
Google DeepMind and Google Labs began rolling out a new Project Genie experience to eligible Google AI Ultra subscribers on November 18, 2026. The update gives those subscribers first-time access to Genie 4, alongside an upgraded Genie 3, a conversational build agent called Project Genie agent, and a community gallery. Project Genie now brings together Google’s most advanced world models and Gemini models so users can generate, step inside, and shape worlds directly from a browser.
What Genie 4 changes
Google DeepMind describes Genie 4 as its most realistic and interactive world model yet. The vendor reports that a user can type new prompts while navigating a generated world and alter weather, add characters, or redirect action in real time. For the first time, those visuals are paired with real-time native audio that adapts as scenes unfold, from ambient soundscapes to character dialogue. The announcement includes no third-party benchmark scores or quantitative latency figures.
Genie 3 remains available in the model picker when users want longer runs or panorama-seeded spaces. Today’s update extends Genie 3 sessions to up to three minutes per generation, up from 60 seconds, and adds world chaining, 360-degree panorama image grounding, and real-time in-world events prompted in natural language.
| Capability | Genie 4 | Genie 3 (updated) |
|---|---|---|
| Visual/audio profile | Highest-fidelity visuals, physics, native audio (vendor-reported) | Lower fidelity than Genie 4; no native audio called out |
| Session length | Not specified in announcement | Up to 3 minutes per generation, up from 60 seconds |
| World chaining | Not specified | Links multiple worlds into continuous journeys |
| Grounding | Not specified | Seeds from 360-degree panorama images |
| Real-time editing | Prompt mid-session; weather, characters, action | Promptable in-world events |
The agent, Gemini observer, and gallery
The Project Genie agent turns natural language into a starting world: it orchestrates Google’s multimodal models to create initial visuals that define a world’s signature look, plus a custom background score matched to the mood. Gemini models can also join an active session as an observer, surfacing context-aware suggestion chips for where to go next or triggering in-world events automatically.
The gallery mixes curated worlds with community work. “Living Memory” transforms personal photos into explorable spaces; “Dino Field Guide” uses 360-degree panoramas of prehistoric habitats; “Rock Paper Scissors” has Gemini acting as a real-time referee. Two named collaborations use Genie 4: “Villa Giallo,” created with creative technologist Will aka BLVCKL!GHT, turns a gothic manor into an interactive room-by-room mystery, and “The Life of Ants,” built with filmmaker Anand Gandhi, draws on Genie 4’s photorealism to navigate an ant colony. Users can remix underlying world structure, soundtrack, or rules and publish their creations to the gallery.
Availability, safety, and known limits
Eligible Google AI Ultra subscribers can try the new Project Genie experience starting today, while Google AI Pro members can join a waitlist on the Project Genie page. All generated video and audio includes SynthID watermarking, and all model inputs and outputs pass through safety screenings.
Google DeepMind states that Genie 3 and Genie 4 remain frontier research world models. It names four challenges still being tackled: camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence. A full Genie 4 model research overview is referenced for deeper limitations. The post does not disclose pricing, API availability, or compute details.
AI Mastery analysis
Genie 4 is a product release, not a model card. The shipping artifact is a browser-resident orchestration surface across Genie 4, Genie 3, Gemini, a build agent, and a gallery, which fits the pattern that deployment-layer interfaces, not generated text, are becoming the product. Real-time promptable editing and adaptive audio push world generation from offline video synthesis toward interactive streaming, where latency and scene permanence become binding constraints. The fact that Genie 3 is capped at three minutes per generation, while Genie 4’s session length is unspecified, suggests compute or state-retention limits still dominate.
The absence of independent benchmarks and reliance on curated demos also makes Genie 4 a case of gated capability verification: capability claims are visible, but third-party verification paths are not. Genie 4’s move from experimental prototype to an Ultra-tier feature marks world models as a consumer-adjacent surface rather than only a research artifact. Watermarking, safety screening, and a curated gallery suggest Google is testing interactive generation inside controlled distribution before any broader API or third-party training use. For AI builders, the nearer-term consequence may be less the model weights than the orchestration and state-management stack that makes real-time world editing usable at all.
Sources
Frequently asked questions
Does Genie 4 run in the browser?
Yes. Google DeepMind says Project Genie brings together its world models and Gemini models so users can generate, step inside, and shape worlds directly from the browser. The new Genie 4 experience is rolling out to eligible Google AI Ultra subscribers.
How long can Genie 3 sessions run after the update?
Up to three minutes per generation, up from the previous 60 seconds. Users can also link multiple Genie 3 worlds into a longer continuous journey.
Can Google AI Pro members use Genie 4?
Not immediately. Eligible Google AI Ultra subscribers can try the new Project Genie experience today, while Google AI Pro members can join a waitlist on the Project Genie page.
What does the Project Genie agent do?
It takes a natural-language description and orchestrates Google's multimodal models to generate initial visuals establishing a world's signature look and a custom background score matched to the mood. Gemini models can also observe a live session and surface context-aware suggestion chips or trigger in-world events.
Does Project Genie watermark generated video and audio?
Yes. All video and audio generated in Project Genie includes SynthID watermarking, and all model inputs and outputs pass through safety screenings.
Related Reading
Google Claims Gemini 4 Hydrogen Runs 2.5x Faster Than 3.8 Flash
Google DeepMind's Gemini 4 Hydrogen is a fast execution tier priced at $1.50/$7.50 per 1M tokens, with vendor-reported 2.5x speed gains over Gemini 3.8 Flash.
NVIDIA Cosmos 3 Open Models Stream 11-Camera Driving Worlds
Four open Cosmos 3 Nano checkpoints generate synchronized 11-camera driving video, LiDAR, and robot action rollouts under OpenMDW1.1.
Gemini Omni 1.2 Flash: 15-Second Clips, 30% Price Cut
Gemini Omni 1.2 Flash is now generally available with 15-second native clips, multi-turn editing, and a 30% price cut to $0.07 per second at 720p.