A Second of AI Video in the Final Cut Costs $0.186, Not the $0.10 on the Price Sheet

A video model's pricing page promises "about $0.10 per second." A 30-second spot is three dollars, you figure, and then the bill arrives at two to three times that.

There are no hidden fees. The price sheet counts every second generated, but only the seconds that survive selection and editing end up in the finished video. Say you order a scene with hands and get three 4-second takes.

In the first two, the hand "melts" and the camera drifts away; the third works. You paid for 12 seconds, and 4 went into the video. That is how the author of a write-up on Habr arrived at $0.

186 for each second that made it into the video, instead of the $0.10 on the price sheet, in a sample from the article. The author generates short videos and ads through the Gemini API and OpenRouter and keeps a log of every take.

If you make AI video for clients, a log like this keeps your estimates from going into the red.

Three Prices for One Second

A second of video has three prices, and mixing them up is expensive.

The first is the price-sheet price. For Gemini Omni Flash it is about $0.10 per second at 720p (5792 video tokens at $17.50 per million comes to $0.101). For Seedance 2.0 via OpenRouter, going by the author's invoices, it works out to about $0.15 per second with audio.

The second is the paid price. Besides the video, you pay for input (images and the prompt), and with Omni you also pay for the model's reasoning. A 4-second clip costs $0.405 by the price sheet; the author's invoice came to $0.42–0.43.

The third is the accepted price: everything spent on a scene's takes, divided by the seconds that made it into the edit. This is the only one you should build an estimate on. The waste ratio ties the three together: divide paid seconds by accepted seconds. It rises with rejects (a whole take goes in the trash) and with trimming (of 5 seconds, 3 went into the edit). People usually forget about trimming, even though on long clips it eats up more than rejects do.

How to Keep the Log and Do the Math

For Omni, the price of a take is calculated from the token counters in the usage field: input at $1.50 per million, video output at $17.50, and text output, including reasoning, at $9. There is a trap here: if you price all output tokens at the video rate, the cost comes out inflated, because reasoning costs nearly half as much. The field names in usage differ between API versions, so save a raw response without the video once and check what they are called in your version. OpenRouter is simpler: video is generated as a job, and when you poll the status, a finished job comes back with a usage.cost field already in dollars.

Next you need a table with one row per take: scene, model, actual length in seconds, price, how many seconds went into the edit, and the reason for the reject or trim. Use the actual length: Omni does not accept a duration parameter, the prompt sets the length, and the model can return anything from 3 to 10 seconds. A small Python script reads the CSV and prints the ratio and the accepted-second price for each scene.

In the sample of three scenes, 23 seconds were paid for in total and 13 were accepted, for a ratio of 1.77. The scene with hands: three takes, 12 seconds paid, 4 accepted, ratio 3.00, $0.318 per accepted second. The scene with a face passed on the first try, but two seconds had to be cut, so the ratio is 1.67 and the price $0.173. The scene with two characters matched the price sheet: $0.105. Averaged over the whole video, it comes to $0.186, almost double the promised $0.10.

Numbers from Real Invoices and How to Bring Them Down

From the author's three projects, as of October 2026, before taxes:

• smoked fish ad, 6 shots on Veo 3.1 Fast: 24 seconds paid and 24 accepted, ratio 1, about $0.10 per second;
• 8 camera moves on Veo 3.1 Fast: 40 seconds paid, 32 accepted, ratio 1.25, $0.125;
• product ad from a single photo on Seedance 2.0: 27 seconds paid, 15 accepted, ratio 1.8, $0.27.

The author calls the first project lucky: start frames were approved before the video, and the camera move in each clip was simple. The product ad is typical of the expensive case: a client will not forgive a warped label, so a shot with a botched package gets reshot no matter what, and the fixes cost almost as much as the first take. Veo 3.1 has been shut off for developers since October 22. These projects were shot earlier, but Omni Flash costs roughly the same by the price sheet.

What lowers the ratio, in the author's experience:

• approve the start frame before generating video: a 1K image costs about 7 cents, a 4-second clip about 42;
• give each clip one simple movement: "picks up a cup" breaks less often than "picks up a cup, drinks, and looks out the window";
• order length to fit the edit, not with a safety margin;
• stop after three identical takes and change one variable: the prompt, the frame, or the model;
• choose a model by its ratio, not by its price sheet: a model that costs twice as much but is accepted on the first take comes out cheaper than one that needs three takes.

If you have a subscription rather than API access, the logic is the same: divide everything you paid for the month by the credits you actually spent. At $20 for 1000 credits and 500 used, a credit costs 4 cents, not 2.

What to Do

The price sheet is a lower bound: you will pay exactly that only if every take goes into the video in full. In the author's experience, the ratio is 1–1.3 on simple scenes, about 2 on product ads, and about 3 on hands, water, and small objects. So in client estimates, the author budgets as much for redos as for the first take, then recalculates from the log after the first test. The figures come from a small sample by a single author, but the method is simple: the log takes five minutes to fill in after editing, and after a couple of projects it shows which scenes are expensive and where to save.