What a 38.7-second AI ad actually costs
Some links on this site are affiliate links. If you sign up through them we may earn a commission at no extra cost to you.
A five-shot, 38.7-second vertical ad cost $0.51 to generate, which is ¥3.42 at ¥6.72 to the dollar. That works out at ¥0.088 for every second of finished video, against a list price of ¥0.09. We estimated ¥3.40 before submitting the job, so the invoice came in 0.6% higher than the estimate.
That is what this article claims: one number, from one job, taken off the billing record. It is not the cost of an ad, and the gap between the two is what the rest of the article is about.
The job
A 38.7-second vertical ad, 720×1280, for a consumer mobile app. Five shots, generated separately and stitched locally into one file. The audio is the model’s own output; no music or voice-over was mixed in.
The script locked the product, the audience, the selling point and the call to action before generation started, so the shots that followed were generated against a fixed script rather than explored for ideas.
The bill
| Item | Amount |
|---|---|
| Shots generated | 5 |
| Seconds billed | 38 |
| Rate (768P) | ¥0.09 / second |
| Generation total | $0.51 (¥3.42) |
| Final runtime | 38.7 seconds |
| Cost per second of finished video | $0.0131 (¥0.088) |
Nothing else was billed. There is no per-request fee, no minimum spend and no storage charge; the cost is seconds generated. A 40-second job we submitted on an earlier project was rejected outright by the API and cost nothing, because a rejected request is not a generated second.
Replacing a rejected shot costs seven cents
We ran the review loop deliberately on a separate project to find out what a rejection costs.
Four shots were generated for ¥1.80 ($0.27). One was rejected, with a note saying the close-up was not tight enough and a detail needed to read more clearly. We regenerated that shot only, with an amended prompt, for +¥0.45 ($0.067). The project totalled ¥2.25 ($0.33).
So replacing a rejected shot costs $0.067.
That number changes how the work gets done. When a person edits a shot, a rejection means a conversation, a re-edit and usually a day of waiting, so teams avoid them. At seven cents the calculation flips, and the sensible move is to generate several versions of the uncertain shot and keep the best one.
What that does is move the constraint. The question stops being whether we can afford another attempt and becomes whether we can tell which attempt is good, and that is a judgement problem the bill says nothing about.
Why this is not the cost of an ad
The $0.51 covers generation. It does not cover:
- Prompting. Each shot needs a prompt naming subject, motion, scene, composition and camera. Writing one properly takes minutes, not seconds.
- Review. Every shot has to be watched and judged. Some are unusable, and you only find that by looking.
- Stitching and checking. Assembling the segments, keeping the audio, confirming the result plays end to end.
There is a fourth omission, and it is larger than those three together. The $0.51 is the cost of the first assembly of these five shots, not the cost of an ad.
Reuse, the term the invoice leaves out
A generated shot does not have to be used once.
Our production does not write a video from top to bottom. It fills a template, and a template is a row of slots: an A-shot, a screen recording, a B-shot, an end card. Shots are generated in pools by slot type, a hundred A-type and a hundred B-type, and any A can be paired with any B.
That makes output combinatorial rather than linear. Two hundred shots across two slot types give ten thousand possible pairings, and the generation cost of each pairing is already paid. The effect in our own numbers is not marginal. Our cost per finished item is low not because each second of generation is cheap but because assets get reused across finished items.
Human production does not amortise the same way. An editor cutting the hundredth item cuts it from the beginning and is paid for it again, whereas generation work that has already been paid for behaves like inventory.
The two cost structures differ in shape, not just in level:
| Cost as output grows | |
|---|---|
| Human production | Roughly linear; every finished item is paid for again |
| AI production | A function of how many unique assets are needed, then amortised |
So cost per finished item depends on one number: how many finished items can be assembled from one round of generation. We worked that number out separately, with the costs, the yield, and what it does to the comparison with human production: Cost per AI ad is a denominator problem. Until you have it, any single figure for the cost of an AI ad, including the $0.51 above, is a lower bound rather than something to plan against.
There is a ceiling on reuse, and it is not a cost problem: repetition causes audience fatigue, and platforms deprioritise near-duplicate creative.
What the human comparison means
A finished, edited item costs us ¥15–20 in pure labour. On the generation line alone, the ¥3.42 job is 17–23% of that.
We put that comparison in the article in order to take it apart, because it is the one most often made dishonestly:
- ¥15–20 is the price of a finished service, and contains a person’s time end to end.
- ¥3.42 is the price of one component, and contains no one’s time at all.
Side by side without that caveat, the two numbers suggest an end-to-end saving that nobody has demonstrated. The honest version needs our own labour added to the AI side, for scripting, prompting, review and stitching, and we have not costed that yet, so the comparison is still missing its largest term. Whether the output is worth buying at all is a separate question, and we measured that one in a live account.
Nothing here says the output is as good as a human edit either. That is a different question, measured by how the creative performs in delivery rather than by what it costs to make, and this article does not measure it.
Method
- Model: MiniMax H3 via a third-party API channel
- Settings: 768P, 9:16, native audio retained
- Source of costs: the project’s own billing record rather than a published price list
- Measured: August 2026
- Currency: USD primary, CNY as billed. ¥3.42 at ¥6.72 to USD, 14 September 2026
- Human labour figure: our own production cost, ¥15–20 per finished item in pure labour
Model prices and rate limits change within months. If you are reading this long after the date above, re-check the per-second rate before relying on the totals.
What we measure next
Two numbers, in this order.
First-pass acceptance. What share of generated shots are accepted without regeneration? We cannot answer that yet, and the reason is worth stating: our tooling records the cost of every generation to the cent, but of the 14 shots we have generated we recorded an explicit review decision on two. Twelve of the fourteen were never regenerated, which is not the same as being reviewed and accepted, and we would rather not present the first as the second. The measurement is cheap. We simply have not been taking it.
The reuse factor. How many finished items come out of one round of generation before performance decays.
That leaves the question this article deliberately does not answer: whether the output performs. Answering it needs delivery data, not production data.
The one number this article adds is the generation line, $0.51 for five shots and $0.067 to replace a rejected one. What it does not add is the cost of an ad, because that depends on a reuse factor we have not measured.