Google Puts Genie 4 World Generation on Gemini API at $6 a Minute
Google opened Genie 4 world generation via the Gemini API on Nov 19, 2026 at ~$6.05 per minute of video, with 120-second session caps, no persistent 3D scene and paid-tier-only access.
By Olivia Hart
4 min read

What's News
- Google made Genie 4 available through the Gemini API and Google AI Studio on November 19, 2026, one day after the Project Genie subscriber release.
- Effective video pricing is about $0.10 per second, or $6.05 per minute, at 240 video output tokens per frame.
- Sessions cap at 120 seconds; visual coherence is strongest in the first 45-90 seconds and may degrade after 60-90 seconds.
- Closing a browser tab does not stop billing; developers must call complete() or cancel() explicitly.
- Access requires a paid-tier API key; quotas run from 2 concurrent sessions at Tier 1 to 10 at Tier 3.
Google opened Genie 4, its interactive AI world-generation model, to developers through the Gemini API and Google AI Studio on November 19, 2026 — at an effective video rate of roughly $6.05 per minute. The model streams continuous 720p video at 24 frames per second with synchronized native audio, and it responds to keyboard controls and live text prompts.
The release puts a price tag on a technology Google had shown only to consumers a day earlier. The Project Genie update of November 18, 2026 gave eligible Google AI Ultra subscribers their first access to Genie 4 and let Google AI Pro members join a waitlist. That subscriber release focused on browser-based creation, a Project Genie agent and a community gallery. The API adds what developers actually need: sessions, documentation, paid-tier pricing and usage quotas.
What does it cost and how do you build on it?
Genie 4 requires a paid-tier Gemini API key; the free tier does not get access. 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. Because video consumes 240 tokens per frame — 5,760 tokens per second — the listed rates work out to about $0.10 per second. Batch API, Flex inference and Priority inference are unsupported.
Developers call the model at models/genie-4 on the v1beta Interactions API. A world is seeded with text, a single reference image, or both. WASD and arrow keys steer movement and camera orientation, while text events modify the scene or supply character dialogue. Google's documentation distinguishes this responsive frame rendering from offline generation of a fixed video clip.
Sessions are built in three steps: create an interaction, connect over WebRTC or WebSocket, and send one world_setup event before steering with user_action or text input events. Browser apps can use an ephemeral token scoped to a single interaction, keeping the long-lived API key on the backend. Google recommends WebRTC for interactive browser apps; WebSocket serves Python and Node.js workflows, frame capture and custom pipelines.
One operational trap deserves attention. Closing a browser tab does not stop server generation or billing. Developers must call complete() to end generation and billing, or cancel() to abort server-side generation. Explicit shutdown is part of cost control.
How stable are the generated worlds?
Google caps sessions at 120 seconds and identifies the first 45 to 90 seconds as the strongest period for visual coherence and object permanence. The implementation guide warns that visual stability may degrade after 60 to 90 seconds. In plain terms: a full session costs about $12 and stays coherent for barely half of it.
The limits are structural. The guide describes generation over a rolling temporal window without a permanent 3D scene. Look away from an object for eight seconds or more, or turn 360 degrees, and landmarks, doors and buttons can move or reset. For robotics or agent experiments, Google concedes the generated environment cannot be assumed to provide a stable long-term map.
Other constraints are fixed. Output is locked to a 16:9 aspect ratio, with no portrait mode or configurable frame rate, and sessions are served only from US data centers. Keyboard actions get a visual response of about 0.2 to 0.5 seconds at low network latency, versus roughly one to two seconds for mid-session text changes. Round-trip latency above 150 milliseconds noticeably delays steering. The model accepts no live microphone or webcam streams and lacks function calling, Google Search grounding, structured outputs and the ability to resume a completed interaction.
Who can use it, and how much?
Quotas scale with paid project usage tier:
- Tier 1: 2 concurrent active sessions, 50 requests per day
- Tier 2: 5 concurrent active sessions, 250 requests per day
- Tier 3: 10 concurrent active sessions, 1,000 requests per day
Real-time control messages within a session do not consume separate per-minute or per-day request quotas. Google lists prototyping interactive experiences, visualizing spatial concepts, and training robotics and AI agents as possible uses, and says every output carries imperceptible SynthID watermarking.
The takeaway is straightforward: Genie 4 is a programmable environment for brief, controlled simulations, not a hosted persistent world. Applications can generate responsive scenes on demand, but the current design offers no world they can retain and revisit across completed sessions.
Source: unite.ai
More from Olivia Hart
Show full bio
Staff writer covering people, launches and day-to-day practice in entrepreneurship, startups and venture funding at Business Bearings.
2,418 articles