Wan 3.0: 30-Second AI Video With Audio, and What It Really Costs

Correction, 27 August 2026. This post originally said Wan 3.0's document-to-video could not be reached through an API. That was wrong. Document and webpage inputs are fields on the reference-to-video endpoint (file_url and web_url, both requiring enable_thinking), not a separate endpoint, and checking the endpoint list alone missed them. The section below is rewritten rather than quietly deleted, because a review that hides its own mistakes is worth less than one that shows them.

Alibaba put Wan 3.0 into public beta on 6 August 2026 and gave it a formal release on 24 August. We wired all six endpoints into the Studio on 26 August, and the schema had one genuine surprise in it.

Not the length. Not the audio. The price.

The Short Version

Wan 3.0 generates single takes from 2 to 30 seconds, at 480p, 720p or 1080p, with native audio produced in the same pass. It matches the longest generation on our roster and costs roughly a sixth as much to run.

That last point is the one worth your attention, and the rest of this post is the working behind it.

What You Can Actually Run Today

The launch coverage led with document-to-video: feed Wan 3.0 a PDF, a spreadsheet, a slide deck or a web page, and get a video back. It is a genuinely novel input mode and it is the reason the model got written up as widely as it did.

You can, and this post said the opposite for a day.

The API offers three Wan 3.0 endpoints: text-to-video, image-to-video and reference-to-video. When this was first published we reported that as the whole surface, and concluded that document ingestion lived only on Alibaba's own platform. That was wrong, and the correction is more interesting than the original claim.

Document input is not a fourth endpoint. It is two extra fields on the reference-to-video endpoint: file_url takes a document, web_url takes a public webpage, and both require enable_thinking to be switched on. Checking which endpoints existed was the wrong altitude to check at, and it produced a confident, published, incorrect sentence.

The wider lesson survives the correction intact, and arguably lands harder: read the schema, not the launch post, and read the whole schema. We made the same point about Seedance 2.5's phantom 4K, where the marketing page and the endpoint disagreed. Here the marketing page was right and our reading of the endpoint was too shallow.

So if you arrived wanting to turn a pitch deck into a film: you can, through the reference-to-video endpoint, and it is in the Studio. We have written up what that mode is actually good at, and the two places it wastes your money, in turning a PDF into a video.

The Spec, Checked Against the Schema

Verified against the model's API schema on 26 August 2026:

Two details in there are easy to skim past.

Duration is an integer, and it accepts every value. Not a menu of 5, 10 and 15. Any whole number of seconds from 2 to 30, which means you can cut a shot to the length the edit actually needs rather than generating 10 seconds and trimming 3. On a per-second rate that is a real saving, not a convenience.

Audio defaults to on. Wan 3.0 has a plain on/off switch for a synthesised soundtrack, and it starts on. The 2.x line worked differently: Wan 2.6 and 2.7 take an uploaded audio track, and Wan 2.7's text-to-video fills the gap with generated background music when you leave the track out. Anyone porting a Wan 2.7 workflow should know which of those they were relying on. Check before you port. Our console exposes Wan 3.0's toggle, so it is a choice rather than a surprise.

The reference surface is the quiet upgrade. Ten stills, five clips and five audio files feeding one generation is a serious consistency tool, and it is triple what 2.7 accepted.

The Number That Changes the Roster

Here is the comparison that made us re-examine our own model recommendations. Every figure is the live Studio price, published on screen before you press Generate.

Seedance 2.5 has been the long-take answer on our roster since July. A full 30-second single pass at 1080p costs 1,917 credits there. The same length, same resolution, same native audio, on Wan 3.0: 300 credits.

That is 6.4 times cheaper. At 720p the gap is 5.7 times. For where both sit against the rest of the field per second, see what AI video costs per second; for what the retakes add on a long take, the Render Ledger.

I want to be careful about what that does and does not prove. It is a cost fact, not a quality verdict. Seedance 2.5 still accepts a far larger reference set, at 30 images and 10 clips against Wan 3.0's 10 and 5, and for character-continuity work across a sequence that surface genuinely matters. What the number does mean is that the default has moved. If a shot needs to run long and the reference load is modest, starting on Seedance 2.5 now costs you five to six times more than starting on Wan 3.0, and you should have a reason for it.

For most briefs, you will not have one.

Does Prime Earn Its Premium?

Wan 3.0 Prime runs about 40% more per second. Alibaba positions it as the higher-fidelity tier, with better faces and micro-expressions.

We have not put both through a controlled comparison yet, so we are not going to tell you it is worth it. What we can tell you is the arithmetic: Prime at 1080p costs 14 credits per second against the base model's 10. A 30-second take is 420 credits rather than 300. Even at the premium it lands well under a quarter of Seedance 2.5's rate for the same job.

Our working advice, and it is provisional: block the shot on base Wan 3.0, and move to Prime only when the frame is carrying a face in close-up. Wide shots, product moves and establishing work will not repay the difference. When we have run a proper side by side, this section gets the results.

What We Have Not Tested

Standard practice on this blog, and worth stating plainly.

We have verified the schema, verified the rate against the provider's published per-second pricing, wired all six endpoints and generated test renders to confirm the pipeline works end to end. We have not yet run a paying client brief through Wan 3.0, and we have not benchmarked its output quality against Kling or Veo on a controlled prompt set.

So nothing above is a quality verdict. The specs are checked, the prices are real, and the creative judgement is still owed. When we have it, this post gets updated rather than quietly replaced.

Update, 6 October 2026. We have now run base Wan 3.0 at 720p through a controlled set: six shot briefs against seven other models, every render watched and judged. It delivered 3 of the 6 (the dolly in on a face, the product hero and the fabric in motion) at $0.50 a render, tied with Veo 3.1 for the most shots delivered and at a quarter of Veo 3.1's price per render. The full grid is in our 48-render model comparison, and every clip, prompt and cost is in The Shot Ledger. Prime was not part of that test.

One thing we can say from the 2.x line: Wan has historically been weaker than Kling v3 on physical motion and momentum, and stronger on prompt adherence for static or dialogue-led framing. Whether 3.0 closes that gap is exactly the sort of claim we are not going to make before testing it.

Questions People Are Asking

Is Wan 3.0 out yet? Yes. Public beta from 6 August 2026, formal release 24 August. All three generation endpoints run in our studio from 25 credits for a 5-second clip at 720p.

How long can a Wan 3.0 video be? Thirty seconds in a single pass, at any whole number of seconds from 2 upward. That ties Seedance 2.5 for the longest single generation on our roster, at roughly a sixth of the cost.

Can Wan 3.0 turn a document into a video? Yes, and we cover the mode in depth in turning a PDF into a video. Point it at a document or a public webpage and it builds the video from that. It is not a separate endpoint, which is what tripped up the first version of this post: the document and webpage inputs are fields on the reference-to-video endpoint, and both need the model's thinking mode enabled.

What does Wan 3.0 cost? In our studio, 25 credits for 5 seconds at 720p, or 300 credits for a full 30-second take at 1080p, with the price on screen before you commit.

Should I use Wan 3.0 or Seedance 2.5? Wan 3.0 unless you need the bigger reference set. Both reach 30 seconds with native audio; Seedance 2.5 accepts 30 images and 10 clips against Wan's 10 and 5. If your shot needs that much reference material, pay the premium. Otherwise the cheaper model does the same job.

Does Wan 3.0 have audio? Yes, generated natively in the same pass, and it defaults to on. The 2.x line had no such switch (Wan 2.6 and 2.7 take an uploaded track instead), so check the toggle if you are porting an older workflow.

Where This Leaves the Roster

Wan 3.0 does not replace anything so much as it resets the price of a capability we already had. Long takes with sound used to be the expensive end of the menu. They are now the middle of it.

The models that remain genuinely differentiated are the ones doing something Wan 3.0 does not: Seedance 2.5 for heavy reference loads, Veo 3.1 for dialogue realism, MiniMax H3 when the delivery needs 2K or 4K. Everything that was only justified by duration needs a better argument now.

Kling v3 is no longer on that list. It was retired from the Studio in September 2026. It is still worth knowing for physical motion, and our Kling 3.0 review covers it. You just run it on Kling's own platform now.

Wan 3.0 and Wan 3.0 Prime are both live in the Studio today, on non-expiring credits, priced to the credit before you generate. If you would rather hand the brief to someone who runs the full stack daily, that is what our production services are for.