Open Weights DeskWhich video models you can actually download, which you cannot, and what the difference costs.

Renting the open model costs more per clip than calling the closed one

The self-hosting argument for video models usually rests on an assumption nobody checks: that an open-weight model, run on rented GPUs, is cheaper per clip than a current commercial endpoint. For the Wan family in August 2026, it is the opposite, and the gap is not close.

Here are the two numbers.

Route What you get Price
Wan2.2-T2V-A14B on a third-party host (SiliconFlow) one 5-second clip, 480P/720P $0.29
wan3.0-video on Alibaba Model Studio, Beijing one 5-second clip, 480P ¥1.05 ≈ $0.147

Both read from the vendors' own published price lists on 27 August 2026. The second figure is a conversion from RMB at the rate stated on the page; treat the third decimal as approximate.

Calling the current, closed, thirty-second-capable model directly is roughly half the price of renting somebody else's GPU to run the two-generations-old open one.

Why that is not a pricing anomaly

It reads like someone mispriced something. It is not. Three ordinary things produce it.

Scale. The vendor running the model it built, at the volume it runs it, on hardware it bought for that purpose, has a lower unit cost than a marketplace host renting capacity and adding margin. This is the same reason managed databases undercut a VPS running the same engine badly.

Generation. Wan 3.0 is newer, and newer usually means more efficient per output second, not less. The intuition that the fancier model must cost more per second is imported from subscription pricing, where tiers are a packaging decision. Per-second inference pricing tracks compute, and compute per output second has been falling.

Tier structure. Wan 3.0 added a 480P tier that Wan 2.7 and everything before it did not have. The entire 2.7 line starts at 720P. So the cheapest legal way to run Wan 3.0 is a step below the cheapest legal way to run its predecessor, which is a genuinely new option rather than a discount.

That last point is the one I would put in front of anyone optimising a budget on this family. The headline feature of 3.0 is thirty seconds in one pass. The feature that changes day-to-day spend is 480P, because it makes a draft cost a quarter of a final, and drafts are most of what you generate. There is a side-by-side of what actually changed between the two generations if you want the rest of the diff.

The costs the comparison above leaves out

The $0.29 figure is a hosted price — someone else is already absorbing the operational burden. If you run the weights yourself, the comparison gets worse, because you take on:

None of that is an argument against self-hosting. It is an argument against self-hosting for the reason of cost, which is the reason most often given.

Where the real trade is

Strip out the price argument and the honest comparison is short:

Open weights (Wan 2.2) Current API (Wan 3.0)
Data leaves your infrastructure No Yes
Fine-tunable Yes No
Works offline Yes No
Max single-pass length Generation-limited 30 seconds
480P draft tier No Yes
Document / web page input No Yes
Native audio in the same pass No Yes
Cost per clip Higher Lower

Every row above the line is a genuine reason to choose weights, and they are all about control rather than money. Every row below it points the other way.

If your requirement is in the top half, the open model is correct and the price disadvantage is a cost of doing business. If your requirement is in the bottom half, you were never actually shopping for weights — and among the hosted routes, a comparison of where each competing model wins the brief is a more useful thing to read than another licence page. MiniMax H3 is the one worth knowing about if it turns out you do need to run something on your own hardware, and it says so there.

There is also a cheaper way to settle the question than reading about it. The disagreement is almost always about output quality rather than licensing, and that is answerable in one clip: run the closed model in a browser tab and put the result next to whatever your own GPU produced from Wan 2.2. If the gap does not justify the hosting bill, the open weights were the right answer all along and you now know it for the price of a free clip instead of a weekend.