AI Video Playbook

Wan

Alibaba's Wan3.0, a fast and cheap model for clips up to 30 seconds, its reference syntax, and how its API pricing works.

Wan3.0 is the value pick for long single takes. It makes up to 30 seconds per clip at 1080p, it is faster than Seedance 2.5, and it costs 0.10 USD per second at 720p on the API. Its weak spot is audio, so plan to replace the sound. Start with the 30 free seconds on Alibaba Cloud Model Studio (Singapore region).

At a glance

Prices checked on 15 Oct 2026.

ItemDetail
MakerAlibaba
Current versionsWan3.0 (officially launched 24 Aug 2026) in two API models: wan3.0-video and the faster, pricier wan3.0-video-prime. Wan 2.7, 2.6, 2.5 and 2.2 are still on the API.
Open weightsWan3.0 is closed. The newest open version is Wan 2.2 (third party check).
Max duration2 to 30 s, or smart duration (default 5 s). With a video input, input plus output must be 30 s or less.
Resolution480P, 720P, 1080P (default 1080P), MP4 at 30 fps. Ratios 21:9, 16:9, 4:3, 1:1, 3:4, 9:16 or adaptive.
AudioNative dialogue, music and sound effects, on by default. Turning it off does not change the price.
ReferencesUp to 10 images, 5 videos (15 s total), 5 audio clips (15 s total), plus 1 document or 1 web link. Up to 20 references in total. First and last frame mode cannot be mixed with references.
Appwan.video (plan prices not confirmed). Also on Higgsfield, Runway, Magnific and other resellers.
APIAlibaba Cloud Model Studio, Singapore: 0.05 USD/second at 480P, 0.10 at 720P, 0.20 at 1080P (list price, with a limited time 30% discount shown). Other regions: about 0.041, 0.083 and 0.165 USD/second.
Cheapest way in30 free seconds of Wan3.0 on Model Studio (Singapore only, valid 90 days). A 5 s clip at 720P costs 0.50 USD after that.

Third party sites list wan.video plans at 10 USD per month for 300 credits and 40 USD for 1,200 credits. We could not confirm these on wan.video, and one of those sites reported 700 credits for a single 5 s clip, which does not add up. Check the plan page before you buy.

Best for

From community reports:

  • Speed and price. In one organic side by side test, Wan3.0 made a shot in 58 s for 7.5 credits, where Seedance 2.5 took 3 min 15 s and 22.5 credits.
  • Long single takes. Up to 30 s per clip at 1080p 30 fps.
  • Benchmarks. Wan3.0 ranked first on the Artificial Analysis text to video board on 30 Sep 2026.
  • Many inputs at once. It accepts a document or a web link as a reference, which few models do.

Weak at

  • Audio. Alibaba itself says audio texture is still being refined. One organic tester called the audio terrible while praising the visuals.
  • Realism of places. In the same side by side test, environments read as more generated than real, and Seedance won on rain, reflections and camera intent.
  • Prime is not always better. In one run the Prime model was slower and about 50% more expensive.
  • No local option. Wan3.0 is closed, unlike earlier Wan versions.

How to prompt it

References

Call references by type and order, counted separately for each type: "Image 1" (or "Img 1"), "Video 1", "Audio 1". The order is the order you add them.

Timing and multi shot

  • Alibaba's examples split the clip into time ranges such as "(0:00 to 0:03)".
  • For a multi shot text to video, write a storyboard of 4 to 6 seconds per shot, each with a time range.
  • Put dialogue in quotation marks.

Edit and extend

Wan decides the task from your words:

  • Edit: use "edit", "remove", "replace" or "change to".
  • Extend: use "extend" or "continue".

Other settings

  • prompt_extend=true lets Wan rewrite your prompt. Turn it off when you need exact wording.
  • Prompts can be up to 20,000 characters, but shorter prompts are easier to debug.
  • Wan3.0 has no documented negative prompt. The open Wan 2.2 does have a negative prompt field.

Example prompts

A 15 second take in three beats:

Prompt
(0:00 to 0:05) Wide shot. A fisherman in a yellow raincoat pushes a small wooden boat into a misty lake at dawn.
(0:05 to 0:10) Medium shot from the shore, slow pan right. He rows out, ripples spreading behind the boat.
(0:10 to 0:15) Wide, static. The boat becomes a small shape in the mist, a heron takes off in the foreground.
Cold blue light with a thin orange line on the horizon. Sound: oars in water, birds. No music.

A reference to video shot:

Prompt
Image 1 is the woman. Image 2 is the kitchen.
The woman from Image 1 chops green onions in the kitchen from Image 2, then looks up and laughs at someone off screen.
Medium shot, static camera, warm window light from the left.
She says: "Dinner in five minutes."

An edit of an existing clip:

Prompt
Edit Video 1: replace the red car with a white scooter. Keep the street, the people and the camera move unchanged.

An extension:

Prompt
Continue Video 1 for 5 seconds: the dog runs out of frame to the right and the camera holds on the empty garden.

Settings that save credits

  • Draft at 480P. It is half the price of 720P and a quarter of 1080P.
  • Use the free 30 seconds to test style before you add money. Pick the Singapore region, which is the only one with the free quota.
  • Be careful with edit and extend. The input video's seconds are billed too, so editing a 10 s clip into a 10 s result bills 20 s.
  • Stay on the standard model. Prime costs about 40% more per second.
  • Do not turn audio off to save money. It does not change the price. Turn it off only if you will replace the sound anyway.
  • Failed requests are not billed and do not use up the free quota.

Known issues

ProblemWhat to do
Weak or odd audioGenerate visuals, then add voice, music and effects in the edit
Edit runs as a new generationUse an explicit verb: "edit", "remove", "replace", "change to"
References ignored in first and last frame modeThe two modes cannot be combined. Pick one.
Download link expiredOutput URLs are kept for 24 hours

Sources