Three weeks ago, Genie 4 was a demonstration for a chosen few. This week it became something you can actually use: anyone paying for Google’s most expensive AI subscription can step inside one of its worlds, and developers can rent the same capability by the second.

Google rolled out a rebuilt Project Genie experience to eligible Google AI Ultra subscribers on November 18, unlocking first-time access to Genie 4, the interactive world model Google DeepMind introduced in late October as a research preview for trusted testers. A day later, the model arrived on the Gemini API and in Google AI Studio for developers, with Google’s documentation pricing streamed output at roughly 10 cents a second. The update also stretches the older Genie 3 and adds a conversational world-building agent and a community gallery. Google is blunt, though, that both models remain frontier research rather than finished products.

What opened up, and for whom#

Project Genie is the consumer front door: a browser experience where a prompt or an image becomes a navigable world. Until now it ran on Genie 3, with generations capped at 60 seconds; the November 18 update puts Genie 4 in the picker for eligible Ultra subscribers, while AI Pro members join a waitlist.

The developer route opened the next day. Genie 4 is listed as a stable model on the Gemini API, reachable through the Interactions API and AI Studio with a paid-tier API key. It streams continuous 720p worlds at 24 frames per second with synchronized native audio, generating as you move. Google pitches the API for prototyping interactive experiences, visualising spatial designs, and training robots and agents.

A misty prehistoric landscape with tall ferns, a glowing path, and a distant dinosaur silhouette
AI-generated illustration: a prehistoric world of the kind Project Genie’s gallery invites visitors to explore.

The model that edits itself mid-scene#

Genie 4’s headline trick is that the world stays editable while you are inside it. Type a new prompt mid-session and the environment changes in real time: shift the weather, drop in a character, redirect the action. It is the first Genie model to pair those visuals with real-time native audio that adapts as the scene unfolds, from ambient soundscapes to character dialogue.

The rebuilt Project Genie wraps that capability in more scaffolding. A conversational agent turns world creation into a dialogue: describe a vision and it orchestrates Google’s multimodal models to set the world’s look and score. Gemini can also sit inside a live session, suggesting where to go next or triggering events on its own. A new community gallery lets visitors explore and remix curated worlds, from photo-turned-spaces in “Living Memory” to prehistoric habitats in “Dino Field Guide.”

Genie 3 stays in the picker for longer runs and panorama-seeded spaces, with sessions stretched to three minutes per generation, up from 60 seconds, plus world chaining and panorama seeding.

Priced by the second#

For developers, the interesting number is the meter. Google prices Genie 4 in video output tokens — 240 per frame, or 5,760 per second at 24 frames a second — which its documentation translates into approximately $0.10 for each second of streamed output, audio included. A full two-minute session, the maximum length, works out to about $12. The exact per-token rate had not yet appeared on the public Gemini API pricing page when the developer release landed, so treat the figure as Google’s approximate rate rather than a locked tariff.

Genie 4 on the Gemini APISpecification
Video output1280×720 (720p), 16:9, 24 fps
AudioSynchronized native audio, bundled into the video token price
PriceApproximately $0.10 per second of streamed output
Maximum session120 seconds
Time to first frameRoughly 5 to 9 seconds
Steering responseAbout 1 to 2 seconds
Round-trip latencyUnder 150 ms in the US, Western and Central Europe, and East Asia
Serving regionsUnited States, Western Europe, Japan

Sessions run over WebRTC or WebSocket and end at the 120-second cap whether you are done or not. Navigation is keyboard-driven, and text steering takes a second or two to appear.

A glowing transparent sphere containing a miniature mountain landscape with a river and a small town, floating above a laptop
AI-generated illustration: a world model as a machine — a whole landscape rendered live inside a sphere.

The honest limits#

The price of admission is that these worlds are still prototypes, and Google’s own documentation says so in unusual detail. Visual stability may degrade after 60 to 90 seconds inside the 120-second cap. Look away from an object for eight seconds or more, or turn a full circle, and landmarks, doors or buttons can shift or reset. Text, signs and clocks drift over time. The consumer post names the same open problems: camera latency, intricate physical interactions, real-time audio precision, and long-term scene permanence.

Content limits are hard too: prompts or images with identifiable real people, public figures or copyrighted brand IP can be blocked at the start or cut a stream off mid-session. Output is fixed 16:9 landscape, and every frame and sound carries a SynthID watermark.

That candour is the most useful part of the launch: world models have spent two years as sizzle reels, and Genie 4 is the first Google has shipped with a public price, a latency budget and a failure list attached.

What to watch#

The first question is whether the worlds hold up outside the demo reel. Early Ultra subscribers will quickly show whether mid-session steering survives ordinary use or only works on the curated gallery worlds. The second is pricing: whether the approximate rate becomes a published per-token price, and whether Google publishes the session limits its documentation leaves blank. The third is the waitlist — how fast Pro members get in, and whether Genie 4 sessions inside Project Genie run longer than the two-minute API cap. If the answers are good, the next frontier-model announcement may not be a chatbot at all. It may be a place.

Sources: Google, “Genie 4 is now available” (November 18, 2026); PPC Land, “Google charges developers about 10 cents a second for Genie 4 worlds” (November 19, 2026).