Summary: Google DeepMind began rolling out Genie 4 to eligible Google AI Ultra subscribers on November 18, 2026, alongside a rebuilt Project Genie experience in the browser. The new model lets you edit a generated world while you are inside it — type a prompt, and the weather, characters, or action change in real time, now with adaptive native audio. If you are prototyping interactive or immersive experiences, this is worth an afternoon; it is not yet something to build a production dependency on. For background on the world-model line, see our earlier post on Genie 4 and world simulation, and the announcement itself is on the Google blog.
Editing the world while you are still in it
The core change is mid-session prompting. Earlier testers of Project Genie told Google they wanted deeper control over what happens inside generated worlds, and Genie 4 is Google’s answer: you keep navigating your world and type a new prompt, and the environment edits in real time. Google’s examples are shifting the weather, dropping in a character, or redirecting the action.
The other first for Genie 4 is real-time native audio that adapts as the scene unfolds — ambient soundscapes up to character dialogue. That matters for prototyping, because sound is usually the part you bolt on last; here it arrives with the visuals.
Genie 3 stays in the picker for longer runs
Genie 4 gets the fidelity crown, but the older model keeps a job. In Project Genie’s model picker, Genie 3 is the option when you want longer runs or panorama-seeded spaces, and it picked up three updates alongside today’s rollout:
- Sessions extended to 3 minutes per generation, up from 60 seconds — a 3x jump.
- 360-degree panorama image grounding, so you can seed a new world from a panorama instead of a text prompt.
- Real-time in-world events: change the environment with natural-language prompts while exploring, and chain multiple worlds into one continuous journey.
That split is worth reading closely. Google’s own framing — Genie 4 for highest-fidelity visuals, physics, and audio; Genie 3 for duration and panorama seeds — means the longest experience you can generate in one piece is still measured in minutes, whichever model you pick.
Build by conversation, share by gallery
The new Project Genie agent moves creation past a single prompt. You describe your vision in a conversation, and the agent orchestrates Google’s multimodal models to produce the opening visuals and a custom background score matched to the mood. Gemini models can join the session as you explore: they observe the screen, surface context-aware suggestion chips for where to go next, or trigger in-world events on their own as the journey unfolds.
Alongside this, Google launched a community gallery. Curated worlds include “Villa Giallo,” built with creative technologist Will aka BLVCKL!GHT — a room-by-room gothic mystery — and “The Life of Ants,” made with filmmaker Anand Gandhi, which uses Genie 4’s photorealism to let you move through a colony at ant scale. Anyone browsing can remix a gallery world with the agent, tweaking its structure, soundtrack, or rules, then publish back. For a builder, the gallery doubles as a free corpus of what these tools actually produce today.
The limits Google names out loud
Credit where it’s due: the announcement is candid about the frontier. Every video and audio output carries SynthID watermarking, and all model inputs and outputs pass safety screenings. But Google also states outright that Genie 3 and Genie 4 remain frontier research models and lists what it is still working on: camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence. When the vendor names four unresolved problems in the launch post, that is a signal about reliability, not marketing noise.
One practical gap: the sources state no pricing for Google AI Ultra, no regional availability, and no data-retention terms for the worlds you create — all of which matter before you invest real work in the platform.
Our take: If you already pay for Google AI Ultra, act now — the cost of trying is an afternoon, it is fully reversible, and the remixable gallery means you can study finished worlds before making your own. If you are not on Ultra, ignore this for now; Pro members only get a waitlist, so there is nothing to evaluate yet. Either way, wait before depending on it. The limits are Google’s own: camera latency, physics, audio precision, and scene permanence are unresolved, and with a 3-minute-per-generation cap in Genie 3 and no published session limits for Genie 4, you cannot yet budget a finished task or its latency. That is our assumption from the vendor’s own disclosures — no independent latency or reliability numbers exist in the sources. The bar moves from experiment to build if Google publishes solid session and latency figures, or if worlds like “The Life of Ants” demonstrably hold up in long sessions without the permanence problems Google itself flags.
Sources
View the original post on X
