Google opens Genie 4 world model to developers via Gemini API
Developers can now access Google's real-time interactive world model to build games, robotics simulations, and spatial designs.

The short version
- Genie 4 generates continuous 720p video at 24 frames per second with synchronized native audio.
- Users can interact with the generated world using WASD keys or live text prompts to change environment variables.
- The model costs approximately $0.10 per second of streamed video, including stereo audio.
- Google includes an imperceptible SynthID watermark on every output to identify AI-generated content.
Why it matters
This moves world simulation from a closed research project into a usable tool for third-party developers. It provides a foundation for creating interactive environments without needing to build complex physics engines from scratch.
Moving from demo to developer tool
Google has officially opened Genie 4, its latest world model, to the broader developer community. While the model was recently available to Google AI Ultra subscribers as part of Project Genie, it is now accessible through the Gemini API and Google AI Studio. This shift allows builders to move beyond watching demonstrations and start integrating the technology into their own software.
By exposing the model under the identifier models/genie-4, Google is enabling developers to use their own API keys to power interactive experiences. This move transitions the technology from a private subscriber playground into a public-facing infrastructure. Developers can now build third-party applications, games, and agent pipelines using these capabilities.
The release follows a rapid succession of launches from Google DeepMind. By making the model available to the public, Google aims to foster a wider ecosystem of creative tools. This includes applications in filmmaking, gaming, and manufacturing where spatial consistency is a primary requirement.
How the world model functions
Genie 4 functions by taking a text prompt, a reference image, or both to generate a continuous 720p world. It produces video at 24 frames per second while simultaneously generating synchronized native audio. This audio includes everything from ambient soundscapes and specific sound effects to character dialogue.
The system relies on Gemini reasoning to maintain a coherent and consistent world state. For example, if a user types a prompt mid-session, the model can update the weather, lighting, or specific objects in the scene. The underlying architecture uses Gemini Omni to anchor realistic appearances while other Gemini models handle the logic that keeps the simulation stable.
Technical implementation for developers includes support for WebRTC for browser-based apps and WebSockets for server-side code. To ensure security, the system uses single-session ephemeral tokens. This setup keeps the user's main API key off the client side, providing a layer of protection for sensitive credentials.
Interactive controls and performance
Users can navigate the generated environments using standard WASD or arrow keys. The system is designed for responsiveness, with keyboard movements typically landing in about half a second. For more complex changes, text prompts generally take between one and two seconds to reflect in the simulation.
Current sessions are capped at 120 seconds to manage the computational load. The model serves content from data centers located in the United States, Western Europe, and Japan. This infrastructure supports the real-time nature of the model, which is necessary for interactive prototyping and visualization.
Google AI Studio provides an interactive canvas and several starter apps to help developers begin experimenting quickly. These tools are intended to lower the barrier to entry for those who want to test the model's capabilities before building full-scale products. The goal is to provide a sandbox for rapid iteration on spatial designs.
Pricing and safety measures
Accessing Genie 4 requires a paid-tier API key with a specific pricing structure. The costs are broken down by input and output types: $1.50 per million input tokens, $9 per million text output tokens, and $17.50 per million video output tokens. This translates to a roughly $0.10 per second cost for streamed video that includes stereo audio.
To address concerns regarding the provenance of AI-generated content, Google has integrated a safety feature into the model. Every output produced by Genie 4 carries an imperceptible SynthID watermark. This technology helps identify the content as AI-generated while remaining invisible to the human eye.
The pricing and safety guardrails are designed to make the model viable for commercial use. By providing clear costs and a standard way to identify synthetic media, Google is positioning Genie 4 as a professional tool. This allows companies to integrate the model into production pipelines with a clear understanding of the overhead.
The path toward AGI
Google frames the development of world models as a significant step toward achieving Artificial General Intelligence (AGI). By creating environments that can simulate physics and logic, the company believes it can provide safer training environments for robots and AI agents. These simulations allow agents to learn complex behaviors in a controlled, virtual space.
Beyond robotics, the model serves as a creative engine for various industries. Filmmakers can use it to visualize scenes, and manufacturers can use it to prototype spatial designs. It provides a way to generate complex, interactive environments that were previously difficult or expensive to build manually.
The transition of Genie 4 into the Gemini API marks a shift toward democratizing these capabilities. Instead of being a research curiosity, the model is now a building block for the next generation of interactive software. It represents a move toward a world where AI can not only generate static content but also maintain a persistent, interactive reality.
What to watch
- How developers will use the model to train autonomous robotics agents.
- The evolution of SynthID watermarking as more world models are released.
- Whether the 120-second session cap will be extended for long-form gaming applications.
Try it
Check out Google AI Studio to see the interactive canvas and starter apps for Genie 4.
Sources and links
First reported by testingcatalog.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.