Thirty seconds is the headline. It is usually the wrong setting.

Thirty seconds in a single pass is the Wan 3.0 feature everyone leads with, and it is real — not six clips stitched, one generation, 900 frames at 30 fps, nothing to match across a join. Every coverage piece stops there.

What none of them mention is that length is the only parameter that scales your bill linearly while scaling your risk worse than linearly. A thirty-second take that misses is discarded whole. There is no salvaging second fourteen. And the arithmetic on that is unkind: one discarded thirty-second 1080P generation costs the same as twenty-four five-second 480P drafts, or four complete 480P attempts at the same length.

So the useful question is not "can Wan 3.0 do thirty seconds." It can. The question is when to spend them.

The duration rules below — any integer from 2 to 30, -1 for smart duration, and input seconds counting against the same ceiling — are all in the create-task schema, read 2026-08-25.

What the thirty seconds actually is

Wan 3.0 takes any whole number from 2 to 30 seconds. The generation is a single pass, which is a meaningfully different object from a stitched sequence: lighting, grade, performance and the audio bed are decided once, so there is no join to hide and no continuity to rebuild in the edit. For anything with a turn in the middle — a product reveal, a story with a punchline, a camera move that needs room to develop — that is the whole reason the feature exists.

One pass is not the same as one shot, and this trips people up. Left to itself, Wan 3.0 decides whether to cut, and on an unspecified thirty seconds it usually will — often between two or three setups in the first five. That is fine when you want something that looks edited and useless when you need an unbroken take. If you want no cuts, the prompt has to say "one continuous take" in those words. How to control the cuts is its own article.

Two more qualifications. Single-pass describes how it is generated, not a guarantee that nothing drifts; audiences went through the launch demo frame by frame and found continuity slips inside long takes, publicly, within a day. And longer jobs take longer, with a typical generation running one to five minutes and thirty-second clips sitting at the far end of that.

Also worth knowing before you commit: there is no way to cancel a Wan 3.0 job. The published API has no call for it. Once a thirty-second job is submitted you are waiting it out, and with only two concurrent slots, a queued batch of long takes ties up both.

Why long takes fail differently

A five-second Wan 3.0 generation that misses costs you five seconds of compute and one clear lesson. A thirty-second one that misses costs six times that and often gives you a muddier lesson, because more went right and wrong at once and you now have to work out which of eight beats broke.

That asymmetry, priced at Alibaba's published rates:

Read the top row as a budget statement rather than a scare. Four attempts at thirty seconds and 1080P is $24.00, and it is very common to spend that before you have decided what the shot is. The full cost breakdown is here.

The ceiling drops when you feed video in

One Wan 3.0 constraint catches people planning around thirty: input video seconds and output seconds must total thirty or less. Attach a twelve-second reference clip and your maximum output is eighteen. Requests past that are rejected rather than trimmed, and the rule holds even when you let the model pick the length.

Reference images do not count toward this. Neither do audio, documents or links. It is video specifically — which, not coincidentally, is also the only reference type that gets billed.

Smart duration, and what it costs to use

Pass -1 and Wan 3.0 chooses the length from your prompt and materials. It is a good feature for exploration and it has one sharp edge: the price cannot be known before it runs.

Providers handle that in two incompatible ways. Some reserve the thirty-second ceiling and refund the difference once the real length is known; others charge the ceiling and keep it. On a clip the model decides should run six seconds, that is $1.20 against $6.00 at 1080P — the same output, five times the price, decided by a terms page.

Use it when you are genuinely unsure how long an idea wants to be, at 480P where being wrong is cheap. Name an explicit duration for anything you are paying real money for.

Choosing a length by what you are making

The bottom row is the one that saves the most money. Most thirty-second waste in Wan 3.0 is not a wrong length choice, it is a length choice made before the shot existed.

The staged route

Block it short, then commit long:

1. Find the shot. Two or three five-second generations at 480P. You are testing framing, motion and whether the sound works — all of which read correctly at 480P, because it is the same Wan 3.0 model at fewer pixels.
2. Test the hardest beat. If second twenty-two is a face turning to camera, generate that as its own short clip first. Learn whether the model can do it before you buy twenty-nine other seconds around it.
3. Run the full length once, at 480P, with the complete prompt. This is where structural problems surface: drift, pacing, beats that overrun.
4. Finish at 720P or 1080P with the identical prompt.

Four steps, roughly $3.50 to a finished ten-second 1080P clip against $8.00 for the impatient route, and at thirty seconds the gap is $10.50 against $24.00.

Writing a prompt that survives thirty seconds

The technique every credible Wan 3.0 guide converges on is the same one, and it is worth stating because it is not obvious: split the clip into consecutive beats, and give each beat a visible end state.

Not "she walks through the market while the camera follows." That is a mood, and a mood sustains about six seconds before the model starts inventing. Instead: 0–6 s, she steps out of the doorway into the light and stops at the fruit stall; 6–14 s, she picks up a peach and turns it in her hand, camera pushing in to a mid; and so on. The end state is what gives second twenty-five something to be consistent with.

Two numbers to size that against, neither of which the guides give you:

One beat per four to six seconds. A thirty-second clip is five to seven beats. Fewer and the model fills the gaps itself; more and beats start colliding inside a single camera move.

660 characters per second of output. The prompt ceiling is 20,000 characters, so at thirty seconds you have more room than any brief needs. Prompt length costs nothing — Wan 3.0 bills output seconds, not input characters — so vagueness in a long take is a choice, not a constraint. Write the audio direction too; sound is generated in the same pass and costs the same whether you specify it or leave it to chance.

The prompt generator will lay a prompt out in beats if you would rather start from a structure than a blank box.

Where thirty seconds is genuinely irreplaceable

Two cases, and they are worth the money every time.

A continuous camera move that needs to travel. A push from a wide to a close, a walk-and-talk, an orbit around a product. Stitching these means matching motion across a cut, which is the hardest thing to fix in an edit and the easiest to notice.

Anything with a spoken performance that has to breathe. Dialogue with a pause in it, a line landing after a beat of silence. Cut that across two generations and the timing dies. Wan 3.0 generates sound with the picture in one pass, so the whole performance has to sit in one pass too.

If your brief is either of those, stop optimising and buy the thirty seconds. If it is not, the honest default is shorter than you think.

Three questions before you pick a length

1. Does the idea have a turn in it? No turn, no need for thirty seconds.
2. Do you know what the shot is yet? If not, this is a 5-second 480P question.
3. Is there reference video attached? Subtract its seconds from your ceiling.

Questions

How long can a Wan 3.0 video be?

Two to thirty seconds, any whole number, in a single generation. Pass -1 and the model picks. With reference video attached, input and output seconds must total thirty or less.

Is a 30-second Wan 3.0 clip really one continuous take?

It is one generation rather than a stitch, so there is no join to grade across — but the model still chooses whether to cut inside that generation, and left unspecified it usually does. Write "one continuous take" if you want none. Even then, single-pass is not a promise that nothing drifts; long takes can still lose a detail, and viewers do notice.

How much does a 30-second clip cost?

$1.50 at 480P, $3.00 at 720P and $6.00 at 1080P per generation, at Alibaba's published rates. Multiply by the attempts you expect, which is where the real number lives.

Should I use smart duration?

For exploration at 480P, yes. For anything you are paying properly for, name the length — providers either reserve the thirty-second ceiling and refund the difference or simply charge it, and you cannot tell which from the button.

Why did my long clip drift at the end?

Usually too few beats. A thirty-second prompt written as a single mood gives the model nothing to hit after about the sixth second. Five to seven beats, each with a stated end state, is the fix — and test the hardest beat as its own short clip before you buy the full length.

Can I extend a clip past thirty seconds?

Not within one generation. Thirty is the ceiling for a single Wan 3.0 pass, so longer pieces are assembled from multiple generations, with the matching problem that implies at each join.