Google's first world model for developers
Google on Nov. 19 (Nov. 20 Japan time) released "Genie 4," an AI that creates interactive worlds on the fly, in the developer-facing Gemini API and Google AI Studio. The Gemini API release notes describe it as the general availability (GA) of the model "genie-4." Logan Kilpatrick, who works on the Gemini API and AI Studio, introduced it on X as the company's first world model that developers can build with to create new experiences.
Introducing Genie 4 in the Gemini API + @GoogleAIStudio, our first world model for developers to build with and create new world model experiences 🔮 Try it out: https://t.co/hy6wuttqIq
A world model is an AI that mimics how the real world behaves and generates video by predicting the next scene. Google announced Genie 4 on Oct. 27 (Oct. 28 Japan time) and said early access in the Gemini API and AI Studio was "coming soon" (earlier article). The resolution and pricing, which were not disclosed at the announcement, are given in the new materials.
At the announcement, Google also said availability in "Project Genie" on the paid "Google AI Ultra" plan was "coming soon." The latest release notes and Gemini API documentation make no mention of Ultra, and ORIENS could not confirm that it has launched.
720p, 24 fps video and audio, steered by keys and text
Genie 4 builds a world from text, a 16:9 reference image, or both. Video is 1280x720 pixels (720p) at 24 frames per second, with 48 kHz stereo audio generated at the same time. Video and audio are streamed live over WebRTC and WebSocket, communication methods used in browsers and elsewhere.
Users move through the world with the WASD and arrow keys, and can trigger events such as making snow fall by sending text mid-session. According to Google, key presses show up on screen in about 0.2 to 0.5 seconds, while text, which passes through world planning, takes about 1 to 2 seconds.
The model is served from data centers in the United States, Western Europe and Japan. Round-trip latency is under 150 milliseconds in North America, Western and Central Europe, and East Asia; elsewhere, steering is delayed and the frame rate can drop, Google said.
Sessions capped at 120 seconds; objects drift over longer runs
A single session lasts up to 120 seconds. Google says visual coherence and object permanence are strongest during the first 45 to 90 seconds.
Google also lists weaknesses in its guide. Genie 4 does not hold a full 3D scene; it generates the next frames from a rolling window of recent time. As a result, looking away for 8 seconds or more, or turning 360 degrees, can cause landmarks, doors or buttons to shift or reset. Text on signs, gauges and clocks deforms, and object counts drift.
It does not support portrait (9:16) video, changing the frame rate, microphone or camera input, function calling, or grounding with Google Search. Prompts or images containing real people, public figures or copyrighted brands can be blocked at startup or cut off mid-stream.
Video costs about $0.10 per second, more than Omni
Genie 4 is available only on the paid tier of the Gemini API. Per 1 million tokens, the units of data an AI reads and writes, input (text and reference images) costs $1.50, text output $9, and video output, audio included, $17.50. Video is counted at 240 tokens per frame, or 5,760 tokens per second, which Google says works out to about $0.10 per second and about $6.05 per minute. Running a session to the 120-second limit would cost about $12.10 for video output alone, by calculation.
Google's video generation model "Gemini Omni 1.2 Flash" costs $1.50 for input, $9 for text output and $12 for video output, or about $0.07 per second of 720p video. Omni generates clips of 3 to 15 seconds in advance; Google recommends Genie 4 for real-time interactive use and Omni for fixed-length clips.
- Genie 4
- Gemini Omni 1.2 Flash
Besides the free tier, Genie 4 does not support "Batch" for bulk processing, "Flex" for cheaper processing, or "Priority" for prioritized processing. Because closing a browser tab does not immediately stop generation on the server, Google asks developers to send a completion command to stop billing.






