Google DeepMind's Genie 4 world model is now open to developers through the Gemini API and Google AI Studio. Several outlets reported the release on November 19, a day after Google made the model available to Google AI Ultra subscribers through Project Genie. The move turns a research prototype into something developers can build against, and it is priced for paid users only.
What Genie 4 Generates
A developer can start a session from a text description, a reference image or both. The model produces a continuous 720p world at 24 frames per second, with synchronized native audio. Users move through the scene with WASD or arrow keys and type live prompts that change weather, lighting, objects or events while the session runs.
Streaming works over WebRTC for browsers and WebSockets for servers. Google says single-session ephemeral tokens keep the main API key off the client, which matters for any app that runs in a web browser. AI Studio includes an interactive canvas and starter apps for testing the model without much code. Every output carries an invisible SynthID watermark that identifies it as AI-generated.
Pricing and Access
Genie 4 requires a paid tier, and it is excluded from the API's free quota. Google's listed rates are $1.50 per million input tokens, $9 per million text output tokens and $17.50 per million video output tokens. Video output works out to roughly $0.10 per second of streamed content, with stereo audio included. At the full two-minute cap, a single session would cost about $12.
Session length is limited. Developer coverage describes a hard cap of 120 seconds per session. Google's own product page describes typical interactions as lasting around a minute, which suggests the practical window is shorter than the cap. Serving is currently limited to data centers in the United States, Western Europe and Japan.
Limitations Google Acknowledges
Google is candid about where the model struggles. Visual details and object permanence can drift over longer runs, after many prompt edits or while the camera moves through a scene. Keyboard inputs take about half a second to register, and text prompts can take one to two seconds. On-screen text is reliable mainly when it appears in the input description, and real-time audio accuracy is still being improved.
Google also says that modeling interactions among multiple users, multiple agents and shared spaces remains an open research problem. Real-world locations cannot yet be simulated with perfect accuracy. Those gaps matter for anyone hoping to use generated worlds as stand-ins for physical environments.
Google made the model available first to Project Genie users and trusted testers before opening it to developers. That staged path gives the company time to watch cost, quality and misuse before wider access. Developers who want to try it should start in the AI Studio canvas with the starter apps, keep sessions short and check the Gemini API documentation for current model IDs and quotas, since details can change during a preview. Google's prompt guide suggests setting the scene first and then layering live prompts for events, camera angles and sound, which helps keep a session coherent.
Why It Matters
World models are increasingly pitched as training grounds for robots and AI agents. Google highlights robotics and agent training among the model's intended uses. A simulator that responds to language in real time could let teams test behavior before they touch physical hardware, which would lower the cost of early experiments.
The limits are the other half of the story. If objects drift and sessions end after a couple of minutes, generated worlds are better suited to prototyping, visualization and early policy testing than to rigorous evaluation. Teams that want to use Genie 4 for robotics should plan to validate results against real-world trials and treat generated video as a hypothesis generator, not as ground truth.
The pricing also shapes how developers will use it. At roughly ten cents per second, long experiments add up quickly, so teams should end sessions explicitly and cache prompts and outputs where they can. Expect early adoption to come from interactive demos, design previews and research labs with budgets for experimentation.
