Wan 3.0 480P vs 1080P: which tier for which job

You have a job to do and a dropdown with three options in it. Here is the short version; the reasoning is below it.

720P for anything another human will look at. 480P for everything that happens before that. 1080P when the file is the deliverable and detail is the point. In a Wan 3.0 480P vs 1080P decision the price gap is four to one, which is large enough that guessing costs real money over a month.

Pick the tier by what happens to the file next

Not by how good you want it to look — everyone wants it to look good. By who sees it and what they do with it.

The row people get wrong is the fourth one. Sending a client a 480P file to approve saves you very little and invites a conversation about image quality instead of the conversation you wanted about the idea. The three Wan 3.0 resolution tiers run 1:2:4 in price — on this site 8, 16 and 32 credits a second — so a client-facing 720P take costs half of a delivery take, not a quarter of one. Buy the middle tier the moment someone else is watching.

The sound is the same at every tier

This is the part that changes how you budget, and it is easy to miss because the resolution dropdown implies everything scales with it.

Wan 3.0 has no audio tier. resolution controls the picture; audio is a plain on/off switch, and turning it off does not make a clip cheaper. Nothing in the request buys you better sound at 1080P than at 480P.

So the entire audio half of your piece — the dialogue, whether the line fits in the seconds you asked for, whether the music lands on the beat you wanted, whether the room tone suits the space — is testable at a quarter of the price with nothing held back. If you are making anything with speech in it, that is where the savings actually are, and it is a much bigger win than shaving pixels off a picture you are going to re-render anyway.

Two honest edges on that. The sound is generated alongside a picture that will be different at the higher tier, so a music cue timed to a specific gesture can shift when the gesture does. And lip sync is judged on the face, which is the thing a low tier renders worst. Judge the writing and the timing at 480P; judge the mouth at the tier you are delivering at.

What a cheap draft cannot promise you: the same shot

Here is the failure that costs people money. You run a few cheap takes, you find one you like, you rerun it at 1080P with the same prompt — and the camera is somewhere else.

That is not a bug, and it is not you doing it wrong. Three lines in Alibaba's Wan 3.0 API reference explain it, and they are worth knowing before you build a workflow on top of them.

Seeds are a nudge, not a lock. The seed field exists to reproduce a result, and Alibaba's own description of it adds that even with the same seed value, results may not be perfectly identical. Pin it anyway — it narrows the range — but do not plan around it holding.

Your prompt is rewritten before every run. prompt_extend defaults to on, which means a language model rewrites what you typed before Wan 3.0 ever sees it. Two runs of "the same prompt" are two different prompts unless you switch it off. This one surprises people more than the seed does.

The tiers are separate renders. In Wan 3.0, 480P and 1080P are two generations, not one render and an enlargement of it. Nothing carries between them except the instructions you send again.

Put together: a draft settles what you are making. It does not settle the frame. That distinction sounds small until you are the one explaining to a client why the approved version and the delivered version are different takes.

How far apart do they land in practice? That depends on the shot, and anyone who gives you a single number for it — us included — is guessing. What is worth doing is looking at a pair yourself before you commit a budget to the assumption: one prompt, pinned seed, prompt_extend off, an explicit aspect ratio, and one take at each tier. Two short renders will tell you more about your own material than any figure someone else publishes, and if the framing holds for the kind of shot you make, you have just licensed a workflow that costs a quarter as much.

The default tier is the expensive one

Worth its own heading because it costs money quietly: if you do not set resolution, the Wan 3.0 API renders at 1080P.

Most defaults in most products are the safe, cheap option. This one runs the other way. Anything you build against the API should set the tier on every request rather than trusting the fallback — our own Wan 3.0 integration overrides it to 480P for exactly this reason, on the grounds that a bug should cost a quarter of what it could rather than four times.

If your bill has been higher than your arithmetic said it would be, check this before you check anything else.

What drafting actually costs

The clean way to feel the ratio: thirty seconds at 480P costs the same as seven and a half seconds at 1080P. A full-length rough of the whole idea, for the price of one short delivery take.

The break-even is just as simple. Drafting and then delivering costs 8 credits a second plus 32 — 1.25 times going straight to 1080P. So the draft has to prevent roughly one rejected delivery take in four to pay for itself.

If your first full-resolution attempt usually works, drafting is a 25% tax and you should skip it. If you are routinely on your second or third, it pays for itself immediately. Most people are in the second group and assume they are in the first.

One caveat that keeps this honest: the draft only prevents the failures it can detect. Those are prompt failures — wrong action, wrong pacing, a cut you did not ask for. If your rejected takes are usually "the framing is not quite right", a draft will not catch that, and the money is better spent on a second attempt at the tier you are delivering at.

Settings that make a draft worth trusting

A reliable Wan 3.0 draft workflow is mostly about removing the variables you control, so that what changes between runs is the thing you were testing.

1. Turn prompt rewriting off. Otherwise you are testing the rewriter, and it runs again at the higher tier with a different result. The cost is that short prompts genuinely get worse without it, so write the full thing yourself: subject, action, scene, camera, sound.
2. Pin the seed and write it down next to the prompt version. Not a lock, but an unpinned seed guarantees drift.
3. Set the aspect ratio explicitly instead of leaving it adaptive, so both renders aim at the same shape.
4. Change one thing per run. Three drafts moving one variable each teach you three things. Three moving everything teach you nothing.
5. Keep drafts short. Eight seconds usually shows you the subject, the motion and the first beat, which is most of what you are checking. If you are weighing a longer take, thirty seconds is the parameter that scales your bill fastest.

Questions

Is 480P enough for AI video?

For your own eyes, yes. A 480P render from Wan 3.0 shows you motion, staging, pacing and sound faithfully. For anyone else's eyes, no. Send a client 480P and you will spend the call discussing image quality rather than the work. Treat it as the tier you think in and 720P as the lowest tier you hand over.

Will my 1080P version look like the 480P draft?

The idea will carry over; the exact shot will not. The Wan 3.0 tiers are separate generations, a pinned seed is not a guarantee, and prompt rewriting is on by default. Expect subject, action, pacing and audio to survive, and expect framing and fine detail to move.

What resolution does Wan 3.0 use if I do not pick one?

1080P, the most expensive of the three. It is the one default here that is not the cautious choice, and it is worth setting explicitly on every request.

Should I draft at 480P or 720P?

480P for almost everything, because Wan 3.0 answers the same questions about pacing, cutting and sound at either tier, and one of them costs half as much. Step up to 720P only when the thing you are unsure about needs the detail to judge — whether a face reads at distance, whether a logo holds together.

The short version

Treat the Wan 3.0 480P vs 1080P choice as a question about your audience rather than your standards. 480P is for you, 720P is for other people, 1080P is for the file you hand over. Draft cheaply and confidently on everything except the frame itself, take the sound seriously at the low tier because it costs you nothing to, set the resolution explicitly so the default cannot bill you, and skip drafting entirely on the jobs where your first take usually lands.

The pricing page will quote a length at all three tiers side by side if you want the number before you commit to one, and what a clip actually costs covers the parts of a bill that resolution does not explain.