Google opened Genie 4 to developers on November 19, 2026, giving creative teams a way to build live, navigable visual prototypes. A minute of generated video costs about US$6.05, including audio, plus input and text-output charges. To stop billing, developers must explicitly complete the session; closing the browser tab does not immediately stop generation.
A world model generates a scene’s next moments in response to someone’s actions. Google’s developer announcement describes steering with keyboard controls and live text prompts. Our October 27 Genie 4 preview explains the earlier creative capabilities and restricted access; the new decision is how to budget and control a paid API experiment.
What does a one-minute prototype cost?
Google’s Standard pricing lists $17.50 per million video-output tokens, $1.50 per million input tokens and $9 per million text-output tokens. Tokens are billing units: video uses 240 per frame, at 24 frames per second.
Calculated video charge for one minute: 60 × 24 × 240 × $17.50 ÷ 1,000,000 = $6.048, rounded to $6.05.
Emitted video duration | Calculated video charge (USD; audio included) |
|---|---|
15 seconds | $1.51 |
30 seconds | $3.02 |
60 seconds | $6.05 |
120 seconds | $12.10 |
These estimates count emitted video, not elapsed time from pressing Start. They exclude input, text output, taxes, currency conversion, hosting, development and additional attempts. The two-minute figure assumes a full stream is permitted by the account’s limits.
Ten separately generated one-minute demonstrations therefore total $60.48 in video charges, calculated before rounding. Reusing a prompt does not remove those charges: the model specification lists neither context caching nor Batch API support.
There is no free Genie 4 API tier. Google’s billing setup may require a minimum $5 prepayment, or offer eligible accounts Postpay. Five dollars does not cover a full minute’s video charge, even before other usage.
How do you stop generation and billing?
Google’s session lifecycle guide specifies await session.complete() to end generation and billing. The server then sends interaction.completed. By contrast, detach() or leaving a Python connection block disconnects locally without ending the server interaction.
For a prototype, connect the visible Stop control to completion and check for the terminal event. Set a deliberate duration limit and plan cleanup for abandoned clients. A disappearing browser window is not confirmation that the server stopped; the documentation does not specify an exact lost-connection shutdown interval.
Google also offers experimental project spending caps, but warns that billing processing can lag by around ten minutes and allow overages. A monthly cap is therefore not a substitute for ending each session.
Which quotas can interrupt a demonstration?
Google publishes these Genie 4 project quotas:
Tier 1: two concurrent active sessions and 50 requests per day.
Tier 2: five concurrent active sessions and 250 requests per day.
Tier 3: ten concurrent active sessions and 1,000 requests per day.
Live movement and text-control events do not consume separate request quotas. Daily quotas reset at midnight Pacific time. Published capacity is not guaranteed, so check the project’s active limits in AI Studio before scheduling a demonstration.
Concurrency is not a spending allowance. Google also lists account-dependent limits over a rolling ten-minute window: $10 for Tier 1 and $200 for Tiers 2 and 3. Applicability depends on billing history and account standing.
Two full one-minute streams would generate $12.096 in video charges alone—above the Tier 1 spending limit where it applies. That is a budget comparison, not a predicted cutoff: the documentation does not explain exactly how this limit affects an already-running stream.
Can teams outside the US use it?
The model specification says generation runs in US data centres only and warns that round-trip network latency above 150 milliseconds impairs live steering. That is a hosting constraint, not a US-residents-only rule.
Google’s general supported-region list includes India and other non-US countries, but it does not establish Genie access for every account. Verify paid model access and assess latency from the intended users’ location before promising responsive controls.
When is Genie 4 a sensible prototype choice?
Genie 4 requires paid API access, and Google caps each session at 120 seconds. Output is fixed at 720p widescreen video. The model page describes stronger visual consistency during roughly the first 45–90 seconds, not guaranteed stability throughout a session.
Google’s documented limitations include objects shifting when revisited and no resumption of completed interactions. Buying another session therefore does not buy a persistent game world. Generated speech also does not mean live microphone input is supported.
Audience matters too. Google’s API terms require users to be adults and prohibit API clients directed at, or likely to be accessed by, under-18s. Check that condition before planning a public game or classroom demo.
If the deliverable is an editable video clip rather than live navigation, Google’s model guide points to Gemini Omni instead. Our guide to Gemini Omni video lengths and generation costs covers that different workflow; the two products should not be treated as interchangeable.
A bounded first experiment is the practical choice: one 15–30-second scene, one interaction to evaluate and a working Stop control. Budget $1.51–$3.02 in video output per attempt, plus other charges and repeats. If the concept needs an unchanged world that users can revisit later, the documented limits point to a different approach.
