Documentation menu

Guides

What it costs

The rate card, how credit is held and charged, and how to keep the bill down.

In short: you pay per page read, plus vectors and the model's tokens if you switch them on, from prepaid credit. We hold an estimate while it runs and charge what it really cost.

What SKU Price (US$)
Page analysis (layout, tables, key-values) pages.layout $0.0250 / page
Page analysis (read-only OCR) pages.read $0.0040 / page
Swift model input swift.input $2.75 / 1M tokens
Swift model output swift.output $13.75 / 1M tokens
Summit model input summit.input $8.25 / 1M tokens
Summit model output summit.output $41.25 / 1M tokens
Compass embeddings compass.input $0.10 / 1M tokens
Swift input, document cached (first call, 1.25× input) swift.cache_write $3.44 / 1M tokens
Swift input, document read from cache (0.1× input) swift.cache_read $0.28 / 1M tokens
Summit input, document cached (first call, 1.25× input) summit.cache_write $10.31 / 1M tokens
Summit input, document read from cache (0.1× input) summit.cache_read $0.83 / 1M tokens
  1. When you send a document, we estimate its cost and hold that much of your credit. If you do not have enough, you get 402 insufficient_credits and nothing runs.
  2. Once the pages are read and the pictures taken, the hold is adjusted to the real size of the job.
  3. When it finishes (or fails), you are charged only what was used and the rest of the hold is released. A retry after a hiccup is on us, not you.

Amounts in the API are in micro-USD: millionths of a dollar, so 61275 is US$0.061. GET /documents/{id}?include=usage shows one document's charges; GET /usage shows everything. Test keys are always free.

Keep the bill down

  • Build with ak_test_… keys. They cost nothing.
  • Only ask for what you use. With no options you pay for the pages alone; embeddings, extraction and queries each add to the bill.
  • Name just the built-in extraction.fields you need: each one is more for the model to write.
  • Use analysis: "read" (US$0.004 a page instead of US$0.025) when you only need the words, not tables or labelled fields.
  • Use images: false for typed documents with no charts, scans or signatures, when the model reads them.
  • Send the questions you know up front in options.queries, rather than one request each.
  • Check GET /usage?group_by=sku to see where the money goes before you optimise.