Skip to main content
TokenCost logoTokenCost
Model ReleaseNovember 20, 2026·6 min read

Genie 4 is on the Gemini API at about $0.10 a second. A full two-minute world costs $12.10, and closing the tab doesn't stop the meter.

It's the first world model you can buy by the token. The tokens are cheap. There are just 5,760 of them every second.

A game controller on a dark table, its glowing blue light bar reflected on the surface

Photo by Radek Skrzypczak on Unsplash

Google moved models/genie-4 to general availability on the Gemini API on November 19, per the changelog. DeepMind announced Genie 4 on October 27 as a research preview for trusted testers. Now any paid-tier key can open a session, type a prompt or send a reference image, and get back a 720p world that streams at 24 frames a second and reacts to WASD and the arrow keys.

Here is what that costs, in the unit you'll actually think in.

Session lengthVideo tokensCost
10-second test57,600$1.01
30-second demo172,800$3.02
60 seconds345,600$6.05
90 seconds (coherence limit)518,400$9.07
120 seconds (session cap)691,200$12.10

Our arithmetic: 5,760 video tokens a second at $17.50 per million, from Google's pricing page. Prompt and reference-image input is excluded; it's a fraction of a cent.

The price card

Three numbers. Input, meaning your text prompt or a reference image, is $1.50 per million tokens. Text output is $9.00. Video output is $17.50, and the 48 kHz stereo soundtrack rides along inside that rate rather than billing separately.

Video is metered at 240 tokens per frame. At a fixed 24 fps that's 5,760 tokens a second, which Google rounds to about $0.10 a second or $6.05 a minute. There's one resolution, 1280x720, so there's no cheaper 360p mode to fall back to and no 4K upsell.

The list of things it doesn't have is longer. No free tier, no Batch, no Flex, no Priority, no context caching. Function calling, Search grounding and structured output aren't supported either, per the model page. Serving is limited to the United States, Western Europe and Japan.

You've seen this card before

$1.50 in, $9.00 text out, $17.50 video out is the exact card Gemini Omni 1.1 Flash bills today. Omni 1.1 bills 5,792 tokens per second of 720p, Genie 5,760, so per second they land within a tenth of a cent of each other.

Genie 4 (interactive)
$0.1008
Omni 1.1 Flash (until Dec 3)
$0.1014
Omni 1.2 Flash
$0.0695

Per second of 720p video with audio, from the Gemini API pricing page, read November 20, 2026.

The difference is that Omni moved on. Omni 1.2 Flash cut video to $12.00 per million, about $0.07 a second, and from December 3 every Omni 1.1 request routes to it. Genie launched on the old $17.50 rate, so it costs 45% more per second than Omni does today. Veo isn't a comparison on this API anymore: Google's video docs say Veo 3.1 Preview shut down on October 22.

We don't think the premium is unreasonable. Omni renders a 3 to 15 second clip and hands it back. Genie keeps a world running for up to two minutes, takes key presses mid-stream and answers them in 0.2 to 0.5 seconds, per Google's developer guide. That's a harder serving problem, and it's priced like one. It also means Google has room to cut it the way it cut Omni.

The bill doesn't stop when the user leaves

This is the line in the docs most worth reading twice. Google's guide says that exiting a local block or closing a browser tab "detaches the transport without immediately stopping server generation." Ending the session, and the billing, takes an explicit complete() call. cancel() aborts immediately. detach() just drops the socket, and Google says never to rely on it to pause or stop billing.

Google doesn't say how long an orphaned session keeps generating. The session cap is 120 seconds, so our working assumption is the rest of that window: up to about $12 for a player who opens your app, looks for three seconds and closes the tab. If your frontend can disappear without warning, and every web frontend can, put the complete() call on the server side with a timeout, not in a page-unload handler.

The AI Studio playground protects you here: hide the tab for more than 10 seconds and the stream pauses, per the prompting guide. Your own app gets no such guard unless you build it.

Rate limits are your spending cap

Genie has its own rate limits, counted in concurrent sessions and session requests per day. In-session key presses and text events don't count. If every session runs the full two minutes, the daily ceiling looks like this.

TierConcurrentSessions / dayMax video spend / day
FreeNot supportedNot supported-
Tier 1250$604.80
Tier 25250$3,024
Tier 3101,000$12,096

Limits from Google's rate-limit page; spend is our arithmetic at $12.096 per 120-second session.

Tier 1 tops out at 50 sessions a day, which is about 100 minutes of world. That's enough for a prototype and a few demo days, not for a public launch. Two concurrent sessions means your third visitor waits.

The more useful number for a consumer product is per hour of play. Back-to-back sessions cost $362.88 an hour, whatever length you cut them to, because the meter runs per second. A user who plays for 20 minutes a day costs about $121 that day. For comparison, Gemini 3.8 Live charges $0.018 a minute for audio output; a minute of Genie costs 336 times that. Nobody is building a free-to-play game on this yet.

What the money actually buys

Google is upfront about the limits, which helps with budgeting. The model page says visual coherence and object permanence are strongest in the first 45 to 90 seconds. Seconds 90 to 120 cost the same $3.02 as any other 30 seconds but are the weakest part of the session, so capping sessions at 90 seconds is a 25% saving with little loss.

The first frame takes roughly 5 to 9 seconds. Text events mid-session take 1 to 2 seconds to show up. There's no portrait output, no frame-rate setting and no way to resume a session, so a dropped connection means paying for a fresh one. The key_e and space actions are accepted but do nothing on their own.

DeepMind published no benchmarks for Genie 4, and Artificial Analysis had no Genie 4 page when we checked, so we can't give you a cost-per-quality figure the way we do for text models. You'll have to judge the output yourself, which at $0.10 a second is at least cheap to try. A ten-second test is a dollar.

Who should pay for it now

Teams generating training environments for agents or robots, and anyone prototyping game concepts or interactive previews, get a lot out of a $6 minute. A level designer who would otherwise spend a day blocking out a space to test an idea is cheap at that price.

Consumer apps where a stranger can start a session should wait, or at least gate it behind a paid plan. At $362 an hour of play and two concurrent sessions on Tier 1, the economics only work when each session is worth more than $12 to someone. If what you need is a short clip rather than a world you steer, Omni 1.2 Flash does it for 31% less per second.

Questions

How much does the Genie 4 API cost?
$1.50 per million input tokens for text or a reference image, $9.00 per million text output tokens and $17.50 per million video output tokens, with audio included in the video rate. Video bills at 240 tokens a frame, 5,760 a second at 24 fps, so about $0.10 a second or $6.05 a minute. There is no free tier, Batch, Flex, Priority or caching.
What does one Genie 4 session cost?
Sessions run up to 120 seconds. A full one costs about $12.10 in video output. A 60-second session is about $6.05, and 90 seconds, the upper end of the window where Google says coherence is strongest, is about $9.07. The prompt and reference image add well under a cent.
Does closing the tab stop Genie 4 billing?
Not through the API. Google's guide says detaching the connection or closing a browser tab does not immediately stop server generation, and that you must call complete() to end the session and stop billing. Only the AI Studio playground auto-pauses a hidden tab, after 10 seconds.
Is Genie 4 more expensive than Gemini Omni 1.2 Flash?
Per second of 720p video, yes. Genie 4 is about $0.101 a second and Omni 1.2 Flash about $0.070, so Genie costs 45% more. Genie's card matches Omni 1.1 Flash's, which routes to Omni 1.2 from December 3. Omni makes fixed 3 to 15 second clips; Genie streams a world you steer with keys while it renders.

The full card is on the Genie 4 page. For Google's text models, see Gemini 4 Neon and the November price changes, or run your own numbers in the calculator.

Sources