Search for a way to use Wan cheaply and you will eventually find Alibaba's Token Plan: a subscription at ¥39 a month, credits usable across models, with an off-peak rate between 22:00 and 08:00 that drops to a fraction of list. It looks like the answer.
It is not the answer, for four separate reasons, and it is worth writing them down because each one independently disqualifies it — so no amount of working around one of them helps.
Read against Alibaba Cloud's Token Plan terms and model list, 27 August 2026.
The subscription's own wording restricts it to interactive use inside Alibaba's tools and prohibits batch API calling. That is not a rate limit or a fair-use clause you might reasonably operate under; it is the specific activity excluded by name.
If your plan is to point an application at it, the plan is against the terms. This is the first and largest of the four, because it does not matter what the plan contains if you are not permitted to reach it the way you need to.
This is the one that surprises people who got past the first point.
The Wan models included in the plan are wan2.7-image and wan2.7-image-pro — image
models. The video tier inside the plan is a different family entirely (happyhorse-1.1
in text-to-video, image-to-video and reference-to-video variants), not Wan.
So a subscription that appears to cover "Wan" covers Wan for still images. Wan 3.0 video, Wan 2.7 video, and everything else in the video line are outside it.
This is an easy mistake to make from the outside, because the plan's marketing surface says Wan and the model list is a click further in. If you are evaluating it, open the model list.
The plan is limited to one per real-name-verified account holder. There is no path where you scale usage by adding subscriptions, which rules out the workaround that the first two problems would otherwise invite.
Two smaller frictions that show up during integration: the plan's dedicated key does not draw on the new-user free quota, and multimodal calls routed through the plan's text-oriented base URL return a 400. The access shape is different, not just the price.
Setting the plan aside, the honest map of "cheap Wan" is short:
| Route | What it actually is |
|---|---|
| New-user free quota | 30 seconds on wan3.0-video, one-off, 90 days, Beijing region only |
| Token Plan (¥39/mo) | No Wan video models; API batch calling prohibited |
| Pay as you go | ¥0.30 / ¥0.60 / ¥1.20 per output second at 480P / 720P / 1080P in four of five regions |
| Self-hosting | Only possible on Wan 2.2 and earlier — 3.0 has no published weights |
| Third-party hosting of Wan 2.2 | More expensive per clip than calling Wan 3.0, not less |
That last row is counter-intuitive enough that it deserves its own note, and it has one on the companion site about open weights — the short version is that renting someone else's GPU to run an open-weight model from two generations ago costs more per five-second clip than calling the current closed model directly.
There is no long-term, scalable, resellable free path to Wan video. Not through the free quota, which is one-off and regional. Not through the subscription, which excludes both the models and the access method. Not through self-hosting, which is only available on an older generation and is not actually cheaper.
The three levers that do exist are the ones anyone building on this model ends up pulling: which resolution you draft at (480P is a quarter the price of 1080P per second and is new in the 3.0 generation — Wan 2.7 had no 480P tier at all), how many seconds you actually need (two is a legal duration and most tests do not need thirty), and how many attempts you spend before the one you keep.
Anyone advertising unmetered free access to this model is either serving a different model behind the label or paying a bill that will stop.
What a hosted front-end can honestly give away is one clip, and it is worth taking
before you commit a budget to any of this:
the first one on wan-3.run is free — 480P, three seconds, sound
included, on the same wan3.0-video the paid tiers run. Writing the prompt costs
nothing at all, because
the prompt builder has no counter on it; it is
the generation that costs money, which is the whole point of this page.