An allowance denominated in seconds sounds simple until you attach something to a request. Then the question becomes: which seconds count? Here is the whole answer for Wan 3.0, sorted by how surprising it is.
Read against Alibaba Cloud Model Studio billing rules and the API reference, 27 August 2026. The input-billing formula is corroborated by three independent gateway documentations, which is flagged where it is used.
Failed generations. Alibaba's billing rules state plainly that a failed job is not charged and does not consume free quota. Both halves matter. It means you can send the request you are unsure about — an unusual aspect ratio, an aggressive prompt, an asset near a size limit — and if it is rejected, you have lost time and nothing else.
Audio. Speech, sound effects and music are generated in the same pass as the picture.
The audio parameter switches the soundtrack on and off and does not change the
price, because there is no second pass to skip. Turning it off to save money saves
nothing. This is unusual enough in this category to be worth stating twice: at least one
competing model charges roughly 50% more per second when native audio is enabled.
Reference images. Up to ten per request, free.
Reference audio. Up to five tracks, one to fifteen seconds each, free.
Documents. A PPT, PDF, DOC, spreadsheet, Markdown file or Keynote up to 50 pages and 100MB. Free. The model reads it and dramatises what is inside rather than narrating over your slides, and that entire capability adds nothing to the bill.
A public web page. One link per request, free. Alibaba's own wording is that it "only supports parsing publicly accessible web pages that do not require login".
Reference video. This is the exception, and it is the expensive one.
Attached reference video is billed for its own duration at the same per-second rate as your output. The formula three independent gateway documentations state identically:
billable seconds = detected input video seconds + requested output seconds
rounded up to whole seconds. A 4.5-second reference clip plus a 5-second output at 720P is ten billable seconds, not five.
The corroboration for this is in Alibaba's own response schema: the usage object
carries an input_video_duration field alongside output_video_duration. A field that
was always zero and never affected anything would not exist.
So the asymmetry, stated as bluntly as it deserves: ten reference images cost nothing. One fifteen-second reference clip is the most expensive input you can attach. Inside a fixed thirty-second free allowance, a fifteen-second reference video is half your allowance spent before the model renders a frame.
Resolution you did not choose. resolution defaults to 1080P, four times the 480P
rate per second. Against a free allowance counted in seconds this costs you nothing
extra; against a bill it is the single most common way an estimate comes out four times
low.
Length you did not choose. duration accepts -1, meaning the model picks the
length. That is a legitimate creative setting and an awkward billing one, because you
cannot know the number in advance. Reconcile from usage.output_video_duration, which is
the only field that reports what actually got produced.
If you are consuming Wan 3.0 through a reseller rather than through Alibaba directly, ask
how that reseller handles -1 before you rely on it. The documented behaviours differ
sharply: one pre-authorises the full thirty seconds and refunds against actual output;
another charges the full thirty seconds and does not refund. Same parameter, up to
twenty-eight seconds of difference per job.
Two independent 24-hour windows start when a job is submitted. The task_id stops
resolving after 24 hours — a query past that returns UNKNOWN, which means expired,
not failed, and needs its own branch in your code and its own sentence in your UI. And
the returned video_url expires 24 hours after the job, so a finished file that was not
copied to permanent storage is simply gone.
Neither of these costs an allowance second. Both of them can cost you the output you already paid for, which is worse.
| Item | Draws down seconds? |
|---|---|
| Successful generation | Yes — output seconds |
| Reference video attached | Yes — input seconds too, at your output rate |
| Failed generation | No |
| Audio on | No |
| Reference images (≤10) | No |
| Reference audio (≤5 tracks) | No |
| Document (≤50 pages, ≤100MB) | No |
| Public web page link | No |
| Higher resolution | Not against a seconds allowance; 4× against a bill |
If you want to see the same arithmetic applied to a real invoice rather than to an allowance, there is a worked breakdown of the four multipliers between the per-second rate and what you actually pay.
The cheapest way to spend an allowance well is to spend the free half of the work first. Reading a clip back into the prompt behind it draws down nothing, and neither does drafting the text — only the generation does. Thirty seconds is six five-second clips, so the difference between arriving with a prompt and arriving with an idea is most of the allowance.