What a finished AI video clip actually costs

The price on the pricing page is per generation. The price that matters is per usable clip, and those are different numbers because you will generate the same shot several times before one is right. Multiply by your keep rate and the real figure often turns out to be five times what you budgeted.

That is the whole article in one paragraph. The rest is how to work out your own number, and which kind of tool is genuinely for which job.

Why per-generation pricing misleads

Every tool in this category prices per generation, per credit or per second of output. None of them price per clip you actually use, because they cannot — that depends on how fussy you are.

So the arithmetic you need is:

At a keep rate of one in five, a €1 generation produces a €5 clip. At one in ten, €10. Neither number appears anywhere on a pricing page, and the second one is the difference between a workable budget and a surprise.

What keep rate should you expect?

It depends almost entirely on how specific your requirement is:

The pattern is consistent: the more precisely you specify, the lower the keep rate, and the cost curve is steeper than the quality curve. Going from "a city street at night" to "a city street at night, camera tracking left, neon reflecting in a puddle" can quadruple what you spend to get one shot.

The practical consequence is a planning habit rather than a tool choice. Write your shot list, then mark which shots genuinely need to be specific. Usually it is one or two. Let the rest be generic and your budget survives.

How do you measure your own keep rate?

Fifteen minutes, once, and it replaces every estimate in this article with your number:

1. Pick three shots you would actually use — one generic, one specific, one with a character.
2. Generate each until you get something you would genuinely publish, or until you have tried eight times.
3. Count the attempts.

Now you have a real keep rate for your kind of work, and you can price a month honestly: shots per month × attempts per shot × cost per generation.

Do this in a free tier before you subscribe to anything. Most tools in this category offer one, and this is exactly what it is for — see which AI tools are actually free for what those tiers really give you.

Which tool is for which job?

The rankings people publish mostly compare these against each other as if they were the same product. They are not.

Consistency across shots. The hard problem: the same character, recognisable, in six shots. Vidu is built around this. If your work is narrative — a series of clips that belong together — it is the capability to test first, and the one with the lowest keep rate if it goes wrong.

A single strong clip. For one hero shot for an ad or a thumbnail sequence, consistency does not matter and raw quality does. Any of the generators can do this, and the relevant number is cost per finished clip at the volume you publish, rather than per generation — so this is what a free tier is for.

Localising footage you already have. A different job entirely: existing video, another language, matched lips. Wizstar is in this space. Test it on one real product before applying it to a catalogue — translation and lip sync both fail in ways that are obvious to a native speaker and invisible to you.

Bulk generation across models. WaveSpeedAI and BytePlus give you model access priced per run rather than per month. This wins when you are generating a lot, or testing several models against the same prompt. It loses while a subscription still covers your volume.

There is a side-by-side of the first three in which AI video tool.

What these tools are still bad at

Worth being plain about, because it sets expectations that save money:

• Text in the frame. Still unreliable. Add text afterwards in CapCut or Canva.
• Hands and precise physical interaction. Better than it was, still the thing a viewer notices.
• A real product looking exactly like itself. If the packaging has to be right, film it.
• Continuity of small details across shots — the same jacket, the same room. This is what character-consistency tools are for, and even then it is the thing to check.
• Anything a viewer will pause on. Generated video holds up at speed and falls apart under attention.

That last one is the useful rule. AI video is strong as movement behind something else — a background, a transition, an establishing shot under a voiceover. It is weak as the thing being looked at.

What a realistic month costs

For someone publishing weekly, the honest shape is usually:

Everything except the first line is predictable. The first line is the one that ruins budgets, and it is entirely determined by a number you can measure in fifteen minutes and almost nobody does.

The rest of the monthly picture is in what AI tools actually cost per month.