Short answer: The Genie 4 API lets developers stream Google DeepMind’s interactive world model into their own apps through the Gemini Interactions API and Google AI Studio. It generates a steerable 720p world at 24 fps with stereo audio and bills about $0.10 per second of streamed output, or roughly $6.05 a minute, on the paid tier only. The caveat: sessions cap at 120 seconds, worlds can drift after 60 to 90 seconds, and closing a browser tab does not stop the meter.
Key points
- Model ID
genie-4on thev1betaInteractions API, generally available on the paid Gemini API tier. There is no free tier, Batch, Flex or Priority option. - Pricing: $1.50 per million input tokens (text or a reference image), $9.00 per million text output tokens and $17.50 per million video output tokens, with audio included.
- Video is billed at 240 tokens per frame, which is 5,760 tokens per second at 24 fps. That works out to about $0.10 per second and about $12 for a full 120 second session.
- Output is fixed at 1280×720, 16:9, 24 fps, with 48 kHz stereo audio. Users steer with WASD and arrow keys plus text prompts sent mid session.
- Serving runs from the United States, Western Europe and Japan. Google quotes under 150 ms round trip latency in North America, Western and Central Europe and East Asia.
- No function calling, Search grounding, structured outputs, session resumption, portrait output or live microphone and webcam input.
What shipped with the Genie 4 API
Google first put Genie 4 in front of Google AI Ultra subscribers through Project Genie, its browser based world builder, with a gallery for sharing and remixing worlds. A day later it opened the same model to developers. The Hacker News thread on the consumer launch passed 280 points, and the developer release is the part builders can actually ship with.
When we explained Genie 4 as a real-time world model, it was limited to trusted testers and Google had published no pricing, frame rate or latency. Those numbers now exist. You start a world from a text prompt, a single 16:9 reference image, or both. Genie 4 then renders frames live in response to keyboard movement and text events, so a prompt like “press the brew button” changes the scene while the user is inside it. Gemini reasoning runs in the loop to keep the world state coherent, and every output carries SynthID watermarking.
Google pitches three uses: prototyping interactive experiences, visualizing spatial designs, and generating environments to train robots and AI agents. AI Studio ships an interactive canvas and starter apps to try it before writing code.
How a Genie 4 session works
Every session follows four steps in the google-genai Python SDK or @google/genai for JavaScript:
- Create the interaction on your server with
interactions.createandmodel: 'genie-4'. For browser clients, mint a single session ephemeral token withinteractions.authTokens.createso your long lived API key never reaches the client. - Connect with
interactions.connectover WebRTC for browser apps or WebSocket for Python and Node.js. - Seed and steer by sending one
world_setupevent, thenuser_actionkey events and textinputevents. Movement keys respond in about 0.2 to 0.5 seconds. Text edits take about 1 to 2 seconds because they pass through world planning. - Complete the session with
session.complete()orsession.cancel()to stop server side generation and billing.
Two practical details from the docs. WebRTC data channels cap messages at 64 KB, so Google recommends compressing seed images to JPEG under 48 KB, or using WebSocket for larger PNG seeds. And only key_w, key_a, key_s, key_d and the arrow keys have built in behavior. To make E open a door, your client has to catch the key and send a descriptive text event instead.
Genie 4 API pricing: what a session really costs
The meter runs on frames, not requests. At Google’s listed rates, video output alone costs about:
- $0.10 for one second of streamed world
- $6.05 for one minute
- $12 for a maximum length 120 second session
- About $360 for an hour of back to back sessions
Input and text output are billed on top, but they are small next to video. The number that matters is how long a stream stays open. Google’s documentation warns that exiting a local block or closing a browser tab only detaches the transport. Server side generation keeps running, and because Genie 4 has no session resumption, there is no reason to leave it alive. If your app does not call complete() on tab close, idle viewers will keep billing you.
Compare that with offline clips. Google points developers to Gemini Omni 1.2 Flash for 3 to 15 second clips, edits and 1080p or 4K output, and to Gemini 3.8 Live for real time voice agents with tool calling. Genie 4 is the only one of the three that reacts to a user’s movement frame by frame, and it is priced like it.
What it means for builders
- Prototypes and pitches, not products yet. At about $6 a minute per viewer, a public consumer app gets expensive fast. Short, gated sessions for design reviews, real estate or venue walkthroughs, and game concept tests are where the math works today.
- Agent and robotics training. Generated environments are cheaper than building each one by hand, but object permanence is short. Looking away for 8 seconds or more, or turning a full circle, can move or reset doors and landmarks, so do not rely on it for tasks that need a stable map.
- Region matters. Developers and users outside North America, Europe and East Asia should test WebRTC latency first. Higher round trip time delays steering and can lower frame rate. Google suggests placing backend relays near a serving region.
- Content filters can end a stream. Prompts or seed images with identifiable real people, public figures or copyrighted brand IP can be blocked at startup or cut off mid session. Plan a graceful error state.
- Text and counts drift. Signs, gauges and clocks deform, and exact object counts change across camera pans. Do not build UI that depends on readable in-world text.
If you are already routing Google models for coding work, as in our Gemini 4 Neon breakdown, note that Genie 4 is a separate surface. It is not a generateContent model and does not run on the Live API, so it needs its own session handling and budget alerts.
What to watch
- Published concurrent session limits. Google links a rate limits page, but per project concurrency caps were not clear at launch.
- Whether Google adds longer sessions, as it did for Genie 3 in Project Genie, or a cheaper lower resolution tier.
- Independent metered bills that confirm the per second math.
- Whether tool calling or structured outputs arrive, which would let agents read world state instead of only pixels.



