By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
4 min read
Key Takeaways
Google says it opened Genie 4 in the paid Gemini API and Google AI Studio on November 19, a day after the separate Project Genie rollout to eligible Google AI Ultra subscribers. Google lists 720p output at 24 frames per second with audio, up to 120 seconds per session, and about $0.10 per second for video output. Google sells the API; its reported capabilities have not been independently tested here.
What changed today: a research preview became an API
When Google DeepMind introduced Genie 4 on October 27, it was a research preview for trusted testers, with no price, resolution, frame rate or session length published. The November 19 developer post changes that: Genie 4 is now available through the Gemini API, via the Interactions API, and in Google AI Studio, where Google offers an interactive canvas and starter apps. Google lists it as models/genie-4 and requires a paid-tier Gemini API key — there is no free tier.
The consumer side arrived a day earlier. On November 18 Google rolled out a rebuilt Project Genie experience to eligible Google AI Ultra subscribers, giving them first access to Genie 4, upgrades to Genie 3, a Project Genie agent that helps build a world from a conversation, and a community gallery. Google AI Pro members can join a waitlist. The API release is the separate step that lets outside developers build Genie 4 into their own products.
How a session works, from prompt to explorable world
A developer starts a session with a text prompt, a single reference image, or both, sent as a world_setup event through the Interactions API. Google says the world then streams as continuous 720p video at 24 frames per second, with synchronized native audio — ambient soundscapes, sound effects and character dialogue. The session connects over WebRTC for browser apps or WebSocket for Python and Node.js, and browser apps can use a single-session ephemeral token so the main API key stays on the server.
Inside the world, WASD keys move and arrow keys move the camera. Text prompts sent mid-session steer the scene — changing weather, lighting, objects or events — and Google's documentation puts the response to a text change at about one to two seconds, against roughly half a second or less for keyboard movement at low network latency. Google names three uses: prototyping interactive experiences, visualizing designs for spatial concepts, and training robotics and AI agents. Every output carries imperceptible SynthID watermarking.
The meter: about 10 cents a second, two minutes at a time
Genie 4 is billed in tokens, but the meter runs on video frames: 240 video output tokens per frame, or 5,760 tokens per second at 24 frames per second. Google's listed standard pricing is $1.50 per million tokens for text or reference-image input, $9 per million for text output and $17.50 per million for video output with stereo audio included — an effective rate of approximately $0.10 per second, or about $6.05 per minute. At that rate, 120 seconds of video output costs about $12.10, excluding any input and text-output charges. Google sells the API and publishes these rates.
Google lists these limits by paid usage tier; actual project capacity can vary: Tier 1 allows 2 concurrent sessions and 50 requests a day, Tier 2 allows 5 and 250, and Tier 3 allows 10 and 1,000. Cost control is partly manual: closing a browser tab does not immediately stop server generation, so developers end billing by explicitly completing or cancelling the session.
Where it breaks — and why Google still calls it research
The limits are plain in Google's own documentation. Sessions stop at 120 seconds and cannot be resumed. Visual stability may degrade after 60 to 90 seconds; looking away from an object for eight seconds or more, or turning through 360 degrees, can make landmarks, doors or buttons shift or reset, because the model generates over a rolling window rather than holding a permanent 3D scene. Output is fixed at 16:9 and 720p — no portrait format and no configurable frame rate — and prompts or reference images showing identifiable real people, public figures or copyrighted brand IP can be blocked at the start or cut off mid-stream.
Google says the same thing from the consumer side: in the November 18 Project Genie post it calls Genie 3 and Genie 4 'world models in the frontier research space' and lists camera latency, intricate physical interactions, real-time audio precision and long-term scene permanence as open challenges it is still working on. Those specifications and limitations are from Google and have not been independently tested here.