Free Seconds LedgerHow many seconds of each Wan model you actually get for nothing, and where the number comes from.

The free-seconds ladder runs backwards: the older the Wan model, the more you get

Alibaba grants new Model Studio accounts a one-off free allowance per model. The allowance is denominated in seconds of video, and it is not the same size for every model in the Wan family. Read as a table, it runs in the opposite direction to what you would guess.

Model Free allowance
wan3.0-video 30 seconds
wan3.0-video-prime 30 seconds
wan2.7-t2v 50 seconds
wan2.6-t2v 50 seconds
wan2.5-t2v-preview 50 seconds
wan2.2-t2v-plus 50 seconds
wanx2.1-t2v-turbo 200 seconds
wanx2.1-t2v-plus 200 seconds

Read against Alibaba Cloud Model Studio's model-pricing and new-user free-quota pages, 27 August 2026.

The current flagship gives you the least. The two-generations-old model gives you nearly seven times as much. This is not perverse — the older models are cheaper per second to run, so a larger allowance costs the vendor the same — but it inverts the intuition that a free tier scales with how much the vendor wants you to try something.

What thirty seconds buys on Wan 3.0

Six five-second clips. Or one clip at the model's maximum length, once.

That is the honest framing. Thirty seconds is an integration budget, not an evaluation budget. It is enough to confirm that your request builder produces a task ID, that your polling loop terminates, and that a file arrives. It is not enough to compare two prompting approaches, and it is nowhere near enough to decide whether the model suits a production brief.

One detail makes it slightly better than it sounds: each model's allowance is counted separately, and wan3.0-video and wan3.0-video-prime are separate models. So the practical Wan 3.0 allowance is 30 seconds on the standard model plus 30 seconds on Prime, which is also the cheapest way anyone will ever get to compare those two against each other.

They do not pool. You cannot spend Prime's thirty seconds on the standard model, and a dated snapshot version of a model counts as a different model with its own separate allowance.

The three conditions that matter more than the size

It is one-off, not recurring. The allowance runs from the date you activate Model Studio, or the date the model is released, or the date your application is approved — whichever applies — and it lasts 90 days. It does not refresh daily, weekly or monthly. When it is gone it is gone, and the ninety-day clock keeps running whether you use it or not.

It exists in exactly one region. Alibaba's wording is direct: the models listed have free quota only in the China North 2 (Beijing) region, and no free quota in any other region. The price tables for Singapore, Tokyo, Frankfurt and Virginia do not have a free-quota column at all.

This is the condition that catches international developers, because the reasonable first move — pick the region closest to your users — silently forfeits the entire allowance. And the region is not a runtime switch: the model, the endpoint and the API key all have to be in the same region, so choosing wrong means minting a new key in the right one and starting over.

Failures do not consume it. Alibaba's billing rules state that failed generations are not charged and do not draw down free quota. This is genuinely good news and it is worth knowing before you start, because it changes how you should spend the allowance: you can afford to send the request that might be rejected. The expensive mistake is not a failed job, it is a successful job at the wrong resolution.

The one that actually eats the allowance

resolution defaults to 1080P.

Free quota is counted in seconds, not in money, so a 1080P second and a 480P second draw the same single second from your thirty. That means resolution does not affect how fast you burn the allowance — but it does mean there is no reason at all to evaluate at 480P while you are inside it. Spend the free seconds at the resolution you actually intend to ship at, because the allowance will not tell you anything about quality at 1080P if you spent it at 480P.

The reverse is true the moment the allowance runs out, and the switch is abrupt: from that point resolution is a 4× cost multiplier, and 480P becomes the correct place to iterate.

How to spend thirty seconds

If the allowance is an integration budget, spend it like one:

  1. Two seconds — smallest legal duration, explicit resolution, explicit ratio, prompt_extend: false. Confirms the async header, the region-key pairing and the polling loop. Do not skip this to save a second; a failed first call teaches you nothing about the model and costs nothing anyway.
  2. Five seconds with audio — speech, effects and music are generated in the same pass as the picture, and the audio toggle does not change the price. This is the run that tells you whether the model's headline capability is real for your material.
  3. Ten seconds with a reference image — the one that shows whether identity holds across a take.
  4. The rest in one long take — because the thirty-second ceiling in a single continuous pass is the thing that distinguishes this model, and you cannot learn anything about it from five-second clips. If you would rather not spend allowance working out how to phrase a long take, there are eleven complete worked prompts printed beside the clips they produced that you can copy before you spend a second of it.

Note what is missing from that list: attaching a reference video. Input video seconds are billed on top of output seconds, and inside a fixed thirty-second allowance that is a straight subtraction from what you have left to look at. Reference images, audio, documents and web links add nothing. Reference video is the one input that costs you allowance before the model produces a frame.