On November 19, Google officially opened access to its latest world model, Genie4, for paid developers via the Gemini API and Google AI Studio. The model identifier for API calls is models/genie-4. This rollout follows the previous day’s launch of Project Genie for Google AI Ultra subscribers, marking a critical milestone: Genie4 becomes the first real-time world model built by Google DeepMind that third-party developers can invoke directly through programmatic interfaces. With this release, the capability is no longer limited to consumer subscription products or demo-only environments.
Genie4 is built on top of Gemini Omni’s visual rendering stack and inherits the reasoning capabilities of the broader Gemini model family. It generates continuous high-fidelity interactive worlds at 720p resolution and 24 frames per second, taking either text prompts or reference images as input. A key differentiator of this model is its native stereo audio output. Alongside visual frames, it produces synchronized environmental ambience, sound effects, and character dialogue, so the generated scene carries coherent audio context without requiring separate audio model pipelines.
End users can navigate within the simulated environment using WASD or directional keys for real-time movement. The round-trip latency for movement commands sits at approximately 0.5 seconds. Developers can also modify scene parameters dynamically with live text prompts, adjusting weather conditions, lighting schemes, and in-scene events. These prompt updates take between 1 and 2 seconds to propagate. The underlying reasoning engine maintains consistent world state across changes, preserving scene logic, object properties and spatial relationships as users move or modify the environment. To support fast iteration, Google AI Studio has rolled out an interactive canvas and starter sample applications. All assets generated by Genie4 carry subtle SynthID watermarks, which enable provenance tracking for synthetic content.
The technical design of Genie4 prioritizes secure streaming and session management. Short-lived session tokens are used for each individual session to protect access credentials. For media delivery, browser-side clients rely on WebRTC, while server-side communication uses WebSocket connections for low-latency streaming. A hard limit of 120 seconds is applied to each independent session. The service is hosted on Google data centers located in the United States, Western Europe, and Japan, which defines the initial geographic coverage boundary for production workloads.
Google has published clear metered pricing for Genie4 API consumption. The rate is $1.50 per million input tokens, $9.00 per million text output tokens, and $17.50 per million video output tokens. Based on these rates, the blended streaming cost averages roughly $0.10 per second of active session runtime. This predictable pricing structure makes it easier for development teams to forecast spend when building interactive prototypes, spatial visualization tools, and agent training pipelines. For teams that manage multiple model endpoints, unified routing through an API gateway can simplify traffic orchestration, and 4sapi offers one such option to consolidate credential management across different model services.
This commercial launch extends the Genie4 research preview that Google DeepMind first shared in October. Before this API release, world simulation capabilities were largely confined to internal research workflows. By opening the API to external developers, Google brings high-fidelity simulated environments into the mainstream AI application ecosystem. The use cases span multiple verticals. Product and UX teams can build interactive spatial prototypes without heavy 3D engine setup. Robotics researchers can use the model as a low-cost, high-fidelity simulation sandbox to train embodied agents. Developers working toward general artificial intelligence (AGI) can leverage the consistent, controllable virtual worlds to run safety tests and benchmark agent decision-making, without the risks associated with physical hardware testing.
It is important to distinguish Genie4 from conventional video generation models. Most existing video models produce static pre-rendered clips. Genie4 operates as a stateful world simulator. It maintains a persistent simulation state, responds continuously to user input, and synchronizes vision and audio streams in real time. This architecture makes it suited for interactive experiences rather than one-off video generation. The latency profile, session time limits, and cost structure are all optimized for short, interactive sessions rather than long-form video rendering.
Even with the new API access, developers should plan their integration around the service constraints. The 120-second session cap means applications need to implement session resumption logic for longer user experiences. Latency numbers are benchmark figures under ideal network conditions; real-world latency will shift based on user location, network quality, and load on the regional data center. SynthID watermarks are embedded automatically and cannot be removed, which aligns with Google’s content provenance policy for generative media.
The release of Genie4 signals a clear shift in the generative AI landscape. The industry has moved beyond isolated text, image and video generation models toward integrated world models that deliver interactive, multi-modal simulated environments. As more developers build on this foundation, the line between generative media and interactive simulation will continue to blur. For builders focusing on embodied AI, spatial computing and interactive storytelling, Genie4 provides a new primitive that was previously only accessible inside large research labs.




