What does it actually cost to generate an AI video ad?

A five second 720p clip from a current video model costs between $0.50 and $2.50. A still image costs between half a cent and 21 cents. So the raw generation for a six shot product ad is roughly $14, and if you are honest about how many attempts it takes, budget three to four times that.

That last number is the one nobody quotes, and it is the only one that matters. Below are the real per call figures, where they come from, and the single setting that quietly multiplied one of our bills by six.

What one generation actually costs

These are the prices we bill against on fal.ai, checked in August 2026. Video is a five second clip at 720p unless noted; images are one image.

The spread inside one row of that table is larger than most people expect. Seedance 2.5 at 720p costs 393 times what a low quality GPT Image 2 still costs. Picking the wrong tier is not a rounding error.

Images are free, video is the budget

Look at the two halves of the table. Every image model fits in the change under $0.25. Every video model starts at half a dollar and climbs.

That has a practical consequence for how you work. Generating stills is cheap enough to be careless with, and animating is not. So do your exploring in stills: get the composition, the lighting and the product angle right as a keyframe you can look at, and only then spend the dollar to make it move. Iterating in video is the single most expensive habit you can have.

This is why our own tool builds the keyframes first and shows them to you before anything animates. It is not a UX preference, it is the price list.

The setting that multiplied our bill by six

Most video endpoints accept duration: "auto" and let the model decide how long the clip should be. It reads like a convenience. It is a billing decision.

Video pricing here is not per clip, it is per token, and the token count is area times time:

tokens = (height × width × duration × 24) / 1024

Duration is a straight multiplier. Seedance 2.5 generates up to 30 seconds, so auto can legitimately return a 30 second clip:

Same call, same model, same resolution. One word in the request, six times the invoice.

We found this while writing the cost estimator, not while reading the docs. The estimator had been costing auto at five seconds, which meant our own spend guard was reserving $2.36 against a call that could bill $14.18. A guard that guesses low is not a guard. It now reserves the maximum, and the actual charge can only come in under it.

If you take one thing from this page: set your durations explicitly.

What a finished ad costs

A short product ad is around six shots. Using the cheap image model for keyframes and a good video model for the clips:

• 6 keyframes with Nano Banana: $0.23
• 6 clips of 5s at 720p with Seedance 2.5: $14.18
• Total, if every single generation is a keeper: $14.41

Nobody's first attempt is a keeper. Three attempts per shot, which is normal and not pessimistic, puts a finished six shot ad at about $43.

Compare that to the alternative honestly. $43 is far below a studio day, and it is also far above the "pennies" figure that gets thrown around. Both halves of that sentence are true and most articles only print one of them.

Which model should you actually pick

There is no best one, which is why the table above does not have a winner column.

• Cheapest usable video: Wan or Sora at $0.50 for five seconds. Fine for background motion and B roll.
• Best value with sound: Veo 3 Fast at $0.75 including audio. Audio costs extra almost everywhere else.
• Highest fidelity: Seedance 2.5 and Sora Pro, at four to five times the cheap tier. Worth it for the hero shot, wasteful for the fourth cutaway.
• Stills: Nano Banana at $0.039 for exploration, Nano Banana Pro or GPT Image 2 high for the one frame that has to be perfect.

The mix is the point. Using your most expensive model for all six shots is how a $14 ad becomes a $40 one with no visible improvement in the shots nobody looks at.

Questions people ask

Is audio included in these prices?

Usually not. Kling and Veo price audio in explicitly, and turning it off makes them cheaper. Seedance 2.5 charges the same whether you generate audio or not.

Why is a five second clip not exactly five times a one second clip?

Because most of these models only offer fixed durations, and the token formula multiplies by resolution as well as time. Doubling the resolution costs more than doubling the length.

Do reference images cost extra?

Image references, no. Video references, yes, and more than you would expect: some endpoints bill the input video's duration on top of the output's, then apply a discount to the total. Read the endpoint's pricing note before you feed it a 30 second reference.

Are these prices going to hold?

No. Every figure here moved in the last year, almost all of it downward. This page carries the date it was checked for exactly that reason. Treat it as a snapshot and check the provider before you plan a budget around it.

Where these numbers come from

Every figure is the estimate our own cost function produces, and that function was written against each endpoint's published per second rate and checked back against the invoices. We are not neutral: we build Tallo, which generates this kind of ad creative, so we have an interest in you believing it is affordable. That is precisely why the retry multiplier is in this article rather than left out of it.

Tallo is built by Speakwise GmbH, a German company that builds AI voice agents and automation. Tallo is where we put what we learned about generating media at scale.

If you want to try it, it is in private testing and the waitlist is on the home page. More pieces like this are on the blog.