AI document extraction pricing here runs on three model tiers: Small at 1 credit per page, Standard at 4, and Large at 12. The fastest way to cut the bill is to check whether your schema fits Small, because Small and Standard run on the same underlying model, so the four times difference buys schema complexity rather than better extraction.
That one fact settles most cost questions. If your schema is ten or fewer flat fields, paying Standard rates gets you nothing extra on that document.
What the three extraction model tiers actually mean
Tier selection is driven by the shape of your schema, not by document type or page count. The API reads the schema you send and works out the smallest tier that can handle it.
Small covers schemas with 10 or fewer leaf fields, no nested objects, and no array-of-object fields. A flat list of values: vendor, invoice number, date, total.
Standard stretches to 20 total fields, counting nested and array subfields. That is the tier for a schema carrying a line-items array or a grouped address block. It runs on the same model as Small.
Large takes anything beyond that envelope, with no separate field cap. It runs a chunked, plan-driven mode that splits a long document into pieces and processes them across several passes on a larger, more capable model. That extra work is what the 12 credits pay for.
| Small | Standard | Large | |
|---|---|---|---|
| Credits per page | 1 | 4 | 12 |
| Schema limit | 10 leaf fields | 20 total fields | No field cap |
| Nested or array fields | Not allowed | Allowed | Allowed |
| Model | Standard | Standard (same as Small) | Large, multi-pass |
| Processing | Single pass | Single pass | Chunked, multi-pass |
Small versus Standard is a shape choice, not a quality one
Both tiers run the same model, so a document whose schema fits Small gets the same extraction it would get at Standard rates. There is no accuracy tax for choosing the cheaper one when your schema qualifies.
Which makes "pick the bigger tier just in case" an expensive habit. Four credits instead of one, on every page, for the life of the integration.
Take an illustrative five-field receipt schema: merchant, date, subtotal, tax, total. Flat, no arrays, comfortably inside Small. At 100,000 pages that is 100,000 credits rather than 400,000, for the same work by the same model.
Why the cheapest available model was not the pick
The model behind Small and Standard is not the cheapest one on the shelf. We evaluated it against a lower-priced alternative on a 38-case labelled dataset.
The cheaper model lost by roughly 4.5 to 5 points on both strict and graded extraction accuracy. It also returned worse-calibrated confidence scores and ran slower. Calibration matters as much as raw accuracy, since it decides which fields get routed to a human. The accuracy gap was worth more than the unit-cost saving, so the tiers price against the better model.
The tier floor is enforced, so you cannot under-buy by accident
Ask for a tier below what your schema needs and the API returns a 400. It does not run the job anyway at the wrong tier, and it does not quietly upgrade you and bill the difference.
The practical effect is that you can try Small first. If the schema fits, it runs. If it does not, you get a clear rejection before credits are spent instead of a surprise on the statement.
Where Large earns its 12 credits
Small versus Standard is a pricing decision. Small versus Large is not. Large exists because some documents genuinely need multi-pass handling: long reports, dense contracts, statements running dozens of pages of repeating structure.
On those, a single-pass extraction runs out of room and fails quietly. Missing rows, values pulled from the wrong section, a total that belongs to a different table. Do not dodge Large to save credits when the document needs it, because the rework costs more than the twelve credits did.
What a credit is worth
Credits get cheaper as you buy more. The entry package is 4,000 credits for ₱2,000, at ₱0.50 per credit, and that rate holds through 10,000 credits (₱5,000) and 50,000 credits (₱25,000). It falls to ₱0.45 at 500,000 credits (₱225,000), ₱0.40 at 1,000,000 (₱400,000), and ₱0.35 at 2,000,000 (₱700,000). The pricing page carries the full ladder and the credit slider.
Two opt-in add-ons sit on top of the per-page rate. Bounding boxes, which return the visual location of every extracted field, add 4 credits per page to any tier. Auto-analyze, which matches an uploaded document against your saved schemas, costs 1 credit per document.
Watch the bounding box rate in particular. On a Small extraction it turns a 1-credit page into a 5-credit page. That is money well spent when you render highlights or need to audit where a value came from, and pure overhead when nothing downstream reads the coordinates.
How to check which tier your schema needs
Count the leaf fields first. Then look for nested objects and arrays of objects, since those push a schema past Small no matter how few top-level keys it has.
If you are over ten leaf fields, ask whether every one of them is consumed downstream. Schemas grow by accretion. Fields get added during a pilot and nobody removes them, and each survivor past the tenth moves the entire workload onto the 4-credit rate.
When the structure genuinely is nested or has repeating line items, Standard is the honest answer. Our guide to AI data extraction covers designing those schemas so they stay tight, and the published accuracy results break performance down per field rather than as one blended figure.
Document extraction cost rewards knowing your own schema. Small when the shape fits, Standard when the structure requires it, Large when the document itself demands it.