2026-10-30 · VideoFreeTier

Seedance 2.5 API Pricing: the Meter Is Tokens, and at 720p a Second Costs More Than the Subscription

The official Seedance 2.5 API is metered in tokens, and the pricing page prints the per-second consequence next to each per-video price: $0.103 at 480p, $0.231 at 720p, $0.569 at 1080p. Those are the vendor's figures, not this site's conversion. Set them against the rates CapCut prints on its pricing page for the same model — $0.12/sec on Pro, $0.18/sec on Ultra, with no resolution attached — and the metered API is the more expensive route at 720p and 1080p and the cheaper one at 480p.

Everything below was read on 30 October 2026 from ai.byteplus.com as a guest, with no account and no API key, and re-read a second time from the Markdown version of the pricing page that the site serves next to the rendered one. The arithmetic in this page is labelled wherever it appears, because the vendor publishes prices and a formula but not the worked example for every shape of request.

The meter is tokens, and the formula sits under a heading called Billing rules

The video generation section of the pricing page opens with a model table whose columns are not prices per clip. They are token rates: $10.70 for 480p and 720p, $11.7 for 1080p, against the row dreamina-seedance-2-5-260628. The page's pricing notes settle the unit — "Token prices are USD per million tokens" — so the row reads as ten dollars seventy per million tokens, not per video.

The same table has an Offline inference column, and for this model it reads Not supported. That matters for anyone hunting a cheaper route: the language-model section of the same document offers a Flex tab at roughly half the standard rate and a Batch rate below that, and the video table has neither. What the video section offers instead is a single formula, filed under Billing rules, several screens below the price.

Estimated token consumption: (Input video duration + output video duration) × output video width × output video height × output video frame rate / 1024

Two things are missing from that sentence and from the table above it: the frame rate, and the pixel dimensions behind "480p", "720p" and "1080p". Both are recoverable, but only by working backwards from the prices, which is the next section.

Running the page's own numbers

The page publishes three worked examples for this model: a five-second 16:9 clip with no video input, at three resolutions. Feed the formula a 24 fps output at 854×480, 1280×720 and 1920×1080 and it lands on all three prices.

The match is to the cent on all three rows, which is why the frame rate and the frame sizes above are written as an inference rather than a quote. Nothing on the page states 24 fps. The 480p dimensions are not stated either, and 854×480 is the only pair among the plausible 16:9 sizes that reproduces $0.514 rather than $0.520 or $0.510.

The same formula also reproduces the six figures on the page's second tab, the one for requests that carry video input: $0.553 to $2.152 at 480p, $1.244 to $4.838 at 720p, $3.062 to $11.907 at 1080p. Six published numbers, one formula, one frame rate. If you are checking this page's figures against your own invoice, the formula is the thing to trust; the frame rate is the assumption to check first.

What a second costs, resolution by resolution

The page prints the per-second figure beside each per-video price, so the ladder can be read without arithmetic at all. What the ladder shows is that a resolution step is close to a doubling plus a bit.

480p and 720p share one token rate, so the price between them tracks pixel count almost exactly: 720p carries 2.25 times the pixels of 480p and costs 2.24 times as much per second, the difference being rounding in the printed rates. The step from 720p to 1080p is 2.25 times the pixels and 2.46 times the price, because 1080p also moves to the higher token rate.

The one-minute column is our arithmetic on the printed per-second rates, and it is the number worth carrying into a budget conversation, because a minute of finished video is a normal thing to want and $34.14 is not a number most people have in mind when they read "video generation model".

The model's own tutorial says it can generate up to 30 seconds in a single request. At 1080p that is about $17.06 of output at the printed rate, and at 480p about $3.08 — again our arithmetic, not a price the page prints, since every worked example on it is five seconds long.

The comparison the API does not invite: CapCut's printed rate

CapCut's pricing page, read on this site two days earlier, prints a per-second rate for the same model inside its paid tiers: Pro "from $0.12/sec (first month)", Ultra "$0.18/sec", Teams "$0.12/sec per seat". Put the two pages side by side on the only unit they share — dollars per second of output — and the direction of the gap depends entirely on the resolution.

Three caveats belong on that table before anyone quotes it. CapCut's rate carries no resolution, so this is not a like-for-like comparison of output quality; it is a comparison of what each page prints, read literally. Pro's rate carries a "(first month)" qualifier inside the line, and the post-promotion figure is not stated anywhere on that page. And a subscription is not a meter: CapCut's plan bundles credits, a credit cap and the rest of its editor, while the API bills each call at the rate above with no ceiling and no bundle.

Read with those caveats, the finding is narrower than "the API is more expensive" and more useful than "prices vary": at the two resolutions most people ship at, the metered API costs more per second of output than the subscription does, and at 480p it costs less. The assumption that going direct to the vendor's API is the cheap route does not survive contact with this table.

Video input cuts the token rate by 40% and raises the bill

The second tab on the same page prices requests that carry video in, and the token rate drops: $6.40 per million at 480p and 720p, $7.00 at 1080p. Against $10.70 and $11.7 that is a 40% cut, and it is the kind of line that gets read as "video input is cheaper".

The formula says otherwise, because the input duration is added to the output duration before anything is multiplied. A 30-second reference video plus a five-second output is 35 seconds of billable duration, seven times the five-second case, against a rate that fell by 40%.

Both ends of each range are the page's own figures, and the 4.19× column is our arithmetic: it is the same at every resolution because the rate cut and the duration increase are both resolution-independent. The saving is real per token and irrelevant per clip.

There is a floor under the low end. The page's note on minimum token consumption says that for this model, when the input includes video, a request whose estimate falls below the minimum is billed at the minimum. The published range starts at 4 seconds of input, which computes to $0.553; a 2-second input would compute to $0.430 and is not offered. The tutorial's recommendation for reference videos is 4 to 30 seconds, which is consistent with 4 seconds being the practical floor rather than a billing artefact — the page does not say which of the two is doing the work, and the minimum table itself sits on an external spreadsheet that was not read here.

Draft mode is not the discount it sounds like

The billing rules mention Draft mode for this model, and the sentence is worth reading slowly: Draft videos are generated at a fixed 480p, and "the token consumption and token rate are the same as normal 480p videos".

So a draft is not a cheap preview of an expensive render. It is a 480p render with a different label on it, priced at exactly the 480p rate. Against a 1080p output it is 5.5 times cheaper, and that saving comes from the resolution, not from the word draft. Anyone drafting at 480p and then rendering at 1080p pays for both, at $0.514 and $2.843, and the second charge is the one that sets the cost of the finished clip.

What is not billed, and what the estimate is worth

One line in the billing rules is unambiguously good news: "You are only charged for successfully generated videos. No fee is charged if generation fails due to reasons such as content moderation."

That changes the arithmetic for anyone generating with prompts that sit near a moderation boundary. A rejected call costs nothing, so a retry is free at the meter; what a rejection costs is time. The consequence is that the cost of a delivered clip depends on how many attempts it takes, not on how many calls you make — and the page does not publish a success rate, because it could not.

The other line to hold onto is that token consumption is an estimate. "Actual consumption is determined by usage.completion_tokens in the API response." The prices above are therefore what the formula predicts, and the invoice is what the response reports. No call was made here, so the two were not compared, and a gap between them is the first thing to check if a bill disagrees with this page.

The same table answers the "more than 2x" comment

The question this page grew out of came from a comment under a Seedance 2.5 comparison video:

"model is literally insanely expensive , more than 2x than original" — @@Limcacs, YouTube comment, collected 29 October 2026

That is a user's impression, recorded as a report and not as a measurement, and it does not say what "original" refers to. The price table does supply one comparison in the same units, though: Seedance 2.5 at $10.70 per million tokens for 480p and 720p against Seedance 1.0 Pro at $2.5 and 1.0 Pro Fast at $1, on the same page, under the same formula.

On that basis a 480p five-second clip is $0.514 on 2.5 against $0.12 on 1.0 Pro — 4.3 times as much, most of it the token rate and a little of it the frame size, since 1.0 Pro's published example uses 864×480. The comment's "more than 2x" is directionally right and understates the table's own spread at 480p. Whether it is right about whatever the commenter was actually comparing is not something this page can settle.

What this page does not publish

The gaps are worth listing, because a pricing page that prints a formula invites the assumption that everything is computable from it.

Every worked example is 16:9 at five seconds. There is no published per-second price for 4:3, 1:1 or 21:9 output, none for durations other than five seconds, and none for the 4K column, which reads Not supported for this model. The page points at a Seedance 2.5 pricing calculator hosted as a Lark form for anything outside those examples; it was not used here, and its answers are not reproduced on this page.

The frame rate is not printed, and neither are the pixel dimensions behind each resolution label. Both were inferred above by reproducing six published figures, which is strong evidence and still an inference.

The page carries no visible revision date. What is written here is the state of the document on 30 October 2026, and the two cells worth re-checking are the token rates in the model table and the three per-second figures in the worked examples.

What this does not settle

No account was created, no key generated and no call made, so the invoice line for a real Seedance 2.5 request was not observed. Whether usage.completion_tokens lands above or below the formula's estimate on a typical prompt is therefore unknown here, and it is the number that decides whether the estimates above are conservative or optimistic.

Whether the offline and Flex routes are genuinely unavailable for this model, as the table's Not supported and its missing Flex tab imply, was read and not tested. And the comparison against CapCut is a comparison of two printed rates, not of two invoices: nobody has run the same 720p clip through both and put the receipts side by side, and until someone does, the 1.93× above is arithmetic on a page, not a bill.

Related on this site: Seedance 2.5 API free — the free-token question asked of the same page, where the 2M headline sits above a row that holds a slash. Seedance 2.5 credits to dollars — the other billing chain, where the unit is a credit and the channel is Invideo; the numbers there are not interchangeable with the token figures here. CapCut Seedance 2.5 free — where the $0.12 and $0.18 per-second rates come from, and which tier's drawer prints them.