Models

Genie 4 API: Budget a Live World Session

Google has priced Genie 4's live API. Calculate video output at $0.1008 per second, then account for inputs and stop each session correctly.

Illustrative miniature landscape inside a glass cube beside a film strip and hourglass

Google made Genie 4 available to developers through the Gemini API on November 19, 2026, adding a paid route to generate and steer a live world. At Google’s listed rates, video output alone works out to $0.1008 per generated second, or about $6.05 for 60 seconds. Text and reference-image input can add to that bill. A prototype budget therefore needs both a duration cap and an explicit end to each session.

What changed on November 19

Google’s developer announcement adds Genie 4 to the Gemini API through the Interactions API and AI Studio. The API produces a steerable 1280×720 world at 24 frames per second with synchronized stereo audio, according to the model documentation. The stable model ID is genie-4.

This is a different access path from the November 18 Project Genie announcement, which discussed consumer availability for eligible Google AI Ultra subscribers and a Pro waitlist. The API documentation calls for a paid-tier Gemini API key. Do not assume a consumer subscription pays for API usage. The DailyTech model tracker gives broader release context; use Google’s newer pricing page for the API rates, since the tracker currently reflects the earlier announcement.

Convert frames into a session budget

Google’s Genie 4 pricing table lists these Standard paid-tier rates. It lists no Free tier, Batch, Flex or Priority option for Genie 4.

Billable component Listed price per million tokens
Text or reference-image input $1.50
Text output $9.00
Video output, including audio $17.50

The interactive-world guide says each video frame consumes 240 output tokens. At 24 frames per second, that is 240 × 24 = 5,760 video tokens per second. Multiply by $17.50 per million tokens: 5,760 ÷ 1,000,000 × $17.50 = $0.1008 per second. This is a calculation from published rates, not a measured invoice.

Illustrative generated duration Video tokens Video-output charge only
30 seconds 172,800 $3.024 (about $3.02)
60 seconds 345,600 $6.048 (about $6.05)
120 seconds 691,200 $12.096 (about $12.10)

For an illustrative pilot of ten separate 120-second sessions, video output alone would be 10 × $12.096 = $120.96. They must be separate sessions because the published session maximum is 120 seconds. These examples exclude text or reference-image input, any text output, taxes, retries and other services. Actual cost depends on billable usage. The AI API cost calculator can help with token arithmetic, but these Genie video-frame calculations should be entered or checked manually; a Genie preset has not been verified.

Closing a tab does not end the bill

The session guide makes an operational distinction: leaving a local block or closing a browser tab detaches the transport, but does not immediately stop generation on Google’s server. It directs developers to call complete() to end generation and billing. A lost connection is therefore a poor substitute for a deliberate termination path.

Google describes four steps: create an interaction, connect via WebRTC or WebSocket, send a world setup and steering events, then complete the session. Browser clients can receive a single-session ephemeral token from a backend; the long-lived API key should stay on the server. For a pilot, design the end control and failure handling before giving users a start control. After a disconnect, check the server-side interaction and your account’s usage rather than assuming the browser event ended the charge. That is a recommended check, not a test we performed.

The same caution applies to any automatic timeout you add. Its purpose is to bound a session; verify that your implementation actually completes or cancels the server interaction. Do not treat a client-side countdown as evidence that billing has stopped.

Check the 120-second ceiling before building

The model page sets a maximum session length of 120 seconds. That makes Genie 4 suitable to evaluate as a short interactive experience, not as one uninterrupted ten-minute world. If your design needs longer play, plan separate interactions and account for the fact that completed interactions cannot be resumed, as the limitations guide explains.

Google also warns that visual stability can degrade beyond roughly 60–90 seconds and that looking away from an object can change it when the view returns. Those are Google’s stated limitations, not results from our testing. The fixed output is 16:9 at 720p and 24 fps; the documentation does not offer a portrait output or selectable frame rate. Test the exact scene and interaction length your product requires before treating a maximum-duration budget as a quality guarantee.

Steering latency matters too. Google says Genie 4 serves from the United States, Western Europe and Japan and warns of higher latency outside its preferred regions. An interactive prototype should be evaluated from its intended users’ regions, with attention to response time as well as cost. For fixed-duration offline video rather than a live world, the guide points readers to Gemini Omni 1.2 Flash; our Omni migration guide discusses that separate workflow. The Runway credit guide offers a different example of budgeting video work, but its credits do not convert into Genie API tokens.

Who should try it now

Genie 4 is worth a small, capped pilot if your product needs an interactive world that reacts to movement or mid-session instructions and can work within a two-minute session. Write down a maximum duration, number of pilot sessions and acceptable total spend first. Then budget input and text output separately from the video estimate, build a reliable complete() path, and inspect billed usage after each session.

If you need long persistent worlds, portrait video, live microphone or webcam input, or a predictable fixed clip, the current documented limitations need to be assessed before integration. Google’s API price makes a trial calculable; it does not establish output quality or the final bill for your workload.

Sources

All pages read on November 20, 2026.

Free toolWhat will 10,000 chats a month cost?Estimate my bill
DailyTech Editorial Team

We explain AI with numbers you can check. Every price, limit and benchmark links to its source and the date we read it. How we work · Report an error