The same AI video model at 7.5× the price
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.
The same video model costs very different amounts depending on where you buy it. We recorded one model at ¥0.11 per second on one channel and ¥0.83 per second on another, a 7.5× spread for identical output. Across ten platforms we tested, the full range is 17×.
The cheap channel is genuinely cheaper, but it is not free. It drops a manual step onto your team, and that step does not appear on the invoice. Channel selection moved our costs more than model selection did, and we only found that out by measuring the manual workaround rather than by reading a price list.
The three channels
“Channel” here means the route you buy access through: a reseller, a consumer app, or the vendor’s own hub. The model underneath is the same one in all three columns.
| Channel type | 768P | 2K | Per minute (2K) |
|---|---|---|---|
| Third-party reseller | ¥0.11/s (≈$0.016) | ¥0.18/s (≈$0.027) | ¥10.80 |
| Consumer app | ¥0.43/s (≈$0.064) | ¥0.70/s (≈$0.104) | ¥42.00 |
| Vendor hub | — | ¥0.83/s (≈$0.124) | ¥49.80 |
For a 38-second advert, the length we typically generate, the gap between the cheapest and most expensive route is ¥27.36 per video (about $4.07). Over 1,000 generations, that is ¥27,360. The model did not change. Only the supplier did.
Why the gap exists
These are not three models. They are three ways of buying one:
- A third-party reseller buys capacity in bulk and passes some of the discount on. Cheapest, least polished.
- A consumer app sells you the interface, the app and the support, and charges separately for watermark removal.
- The vendor hub is closest to first-party pricing.
Nothing about the generated pixels differs. The difference is in everything around them: price, terms, support, and how much of the work you have to do yourself.
What the cheap end actually costs
This is the part that is not in the price column. From our evaluation notes on the cheapest route:
%signs and numerals are sometimes rendered incorrectly on screen. For ad creative with a price or an interest rate in frame, that is a defect rather than a nitpick. This is the same failure mode we measured in image generation, where taking the text away from the model removed the error entirely.- A cross-segment consistency step is manual. Keeping a person looking the same across two stitched segments means re-applying a reference image between them, which is minutes of human time per video.
- It was not slower. We expected the cheap route to be slow; it came in at about 5 minutes per job against 6 minutes on the most expensive one. Our assumption was wrong.
The second point is the one people miss. Saving ¥0.72 per second on a 12-second shot is ¥8.64. If the manual workaround costs ten minutes of an editor’s time, you have spent more than you saved. This is one more reason the per-second rate carries less information than the number you divide it by, which is the argument we make about cost per ad.
We still use the cheaper route, because our volume is high enough that the arithmetic works out. It works out because we measured the workaround. It would not have worked out if we had trusted the price list.
The wider picture
The same effect shows up across the ten platforms we evaluated. Prices below are per second of generated video.
| Platform type | Model generation | Price | Main weakness |
|---|---|---|---|
| Reseller (cheapest) | Latest | ¥0.11/s (768P) · ¥0.18/s (2K) | % and numerals misread; manual consistency step |
| Cloud vendor | wan3-class | ¥0.60/s (720p) | Character consistency does not hold over time |
| Consumer app A | Same as cheapest | ¥0.43/s (768P) · ¥0.70/s (2K) | Watermark unless you buy a membership |
| Vendor hub | Same as cheapest | ¥0.83/s (2K) | 6 min generation; % misread |
| Consumer app B | seedance-class | ¥0.90/s (720P) | Credits easy to waste as a beginner |
| Consumer app C | seedance-class | ¥1.20/s (720p) | Canvas-only workflow, laborious |
| Consumer app D | seedance-class | ~¥1.50/s | Shots sometimes jump to unseen scenes |
| Consumer app E | seedance-class | ~¥1.90/s | Everything requires the infinite canvas |
| Free tier | seedance-class | Free | No audio; blurry; breaks consistency past 10s |
| Consumer app F | — | — | 5–7 day turnaround, 70% failure rate |
From ¥0.11 to ¥1.90 per second, that is a 17× range, which is larger than the quality gap between most of the models listed.
What this article does not claim
- It does not claim the cheap channel is the right choice. It is the right choice for us, at our volume, with the workaround measured. At ten videos a month, it probably is not.
- It does not compare output quality between channels. All three routes run the same model, so we assumed output was identical. We did not run a blind quality test, and we should.
- It does not model commercial terms. Credit packs, prepayment discounts and secondary-market credit resale move effective pricing substantially. One platform above is roughly half price if credits are bought through a resale market, and we have not priced that in.
- It does not claim these prices still hold. They are dated observations.
Two things we are still working on
The model is only one variable, and how people work with it is another. Both have room to improve. This article isolates the supplier, but the manual consistency step it exposes is a workflow problem rather than a model problem. A start-frame image may remove that step entirely, which would make the cheap channel’s advantage structural rather than situational. We set out to test that and have not run the test, so the manual step remains an estimate.
Our sample is one team, one workflow, one month of price records. A single evaluation round establishes that the spread is worth asking about. It does not establish that these are the prices you will see. Prices on this market change within months, and we intend to re-run the survey and publish the movement, including if the spread narrows.
Method
- Source: a structured evaluation of ten AI video platforms, plus our own project billing records.
- Measured: platform prices recorded 2026-08; our own billing figures from 2026-08 and 2026-09.
- Currency: ¥ converted at 6.72 to the dollar (2026-09-14).
- Basis of comparison: published or quoted per-second rates at the time of testing, not negotiated rates.
- Recalled figures: the cost of the manual consistency step is an estimate. We did not time that work, so “minutes of human time per video” is a recollection rather than a measurement. The generation timings, 5 and 6 minutes per job, come from our evaluation records.
A note on our own numbers disagreeing
We have billed the cheapest channel at ¥0.09 per second (768P) and ¥0.15 (2K) in earlier months, against the ¥0.11 / ¥0.18 recorded in this evaluation. Same channel, different dates. Either the rate changed or the credit package did, and we do not currently know which.
We are publishing both figures rather than picking the cleaner one. If your own numbers differ from ours, that is expected on this market, and we would like to hear them.