Every model, priced

Per million tokens, input and output. Official is what the vendor charges you directly. Loot is what we charge. The two are not the same number and that is the entire product.

Callable upstream—

Text models priced—

Text, vs official−5%

Media, vs officialList

Modality Sort

DeepSeek window —

Model	Modality	Loot in / out	Official in / out	vs official
—	—	—	—	—

—

The DeepSeek window

DeepSeek V4 Pro costs half as much upstream outside 00:00–14:00 UTC, and that halving is passed to you rather than kept. The margin on the model is the same in both windows — the price is what moves. Every response carries the window it was billed in and what the same call would cost in the other.

—

DeepSeek V4 Pro vs official

The 20% we buy at is a rate on a dashboard, not a contract. Every price here is a stored cost per model with our margin applied, re-harvested from upstream and diffed rather than typed in — because list prices moved 17–42% in five months while the discount held, and a stale table sells at cost. The official column is the vendor's own published rate, stated here so this page can be checked against it.

The ones we will not sell you

Of — callable models upstream, a number of carry no published price at all. They are not in the catalogue and not callable through this gateway, because a model we cannot price is a model we bill by guessing. Ask for one and the answer is a 409 that says so. The rule is older than the code: never sell a model we cannot price.

Images, video and audio

Media sells at list — official × 1.00 — rather than under it. There is no router undercutting it to answer, and the margin is what pays for generations that fail and for storing the ones that do not. Per-unit prices publish on this page when the endpoints open, alongside the moderation policy and the retention window. Not before: a page whose every read fails is worse than no page.