Document Extraction Pricing — Work Out Your Break-Even in Documents, Not Dollars

Every other cost comparison on this site asks how long until the one-time price pays for itself? Document extraction is the one category where that is the wrong question.

Here the meter runs on documents, not months. So the number that decides your choice is not a date — it is a volume.

Four vendors, four different meters

From each vendor's own pricing page:

These four numbers are not comparable as printed. A page, a credit and a document are different units, and two of the four have floors underneath them.

A correction this re-read forced, and it is about our own sourcing. Earlier versions of this table quoted Parseur at "$39/month" for 100 pages and "$99/mo for 1,000 pages". Those figures came from Parseur's own Nanonets-comparison page, not from its pricing page — first-party domain, wrong document. Read at parseur.com/pricing on 5 September 2026, the pricing page printed no dollar amount anywhere: the paid tiers were described by allowance, the numbers lived behind a slider, and the only rate on the page was "$0 / page" against the free plan. Parseur therefore belonged with Docsumo and Gatekeeper in the column of vendors you cannot budget from, and we had it in the wrong column because we read the wrong page of the right website. Read again on 24 September 2026, the pricing page prints a table as text: Micro, 100 pages, $49 billed monthly or $468 a year billed annually — the $39 a month is that annual price. So Parseur is back in the column you can budget from, sourced to its own pricing page.

The free tier needed a second look too, and a later read the same day settled it in the vendor's favor. We had it as "20 pages/month included permanently." An early reading found a comparison row saying "20+ credits valid for a year" beside another saying "Credits renew every month", and we published that contradiction. *A closer read of the page states it plainly: "Start on our free tier, which includes 20 pages every month." So the tier renews monthly, the guide's original description was right, and the contradiction was ours to resolve rather than the vendor's to answer* (all read 5 September 2026). Recorded rather than quietly deleted, because we published the doubt and should publish its resolution in the same place.

Mindee's card looked as if it did not reconcile with itself, and a later read settled it. On 5 September we recorded $44 a month, a 6,000-credit default, and "≈ 0.044 per credit", and wrote that six thousand credits at 0.044 is $264 — six times the price printed beside it — so we printed all three rather than the one that would make the row tidy. Read again on 26 September 2026, the card itemizes itself: to a US time zone, $53 a month billed annually is a $26 platform fee plus *6,000 credits at "≈ $0.054 each", shown as $27 a month — so the 6,000 credits are a year's allowance, not a month's. Mindee also sets the currency by the reader's time zone: a European one is shown 44 €, 529 € a year* and "≈ 0,044 € each", the same numerals we recorded. Whether the page put a dollar sign beside 44 on 5 September we cannot now tell, so this guide quotes both currencies as they read today, each with its own sign.

Where the floors bite

Veryfi's $500 minimum is the steepest in our whole ledger, and its effect is worth showing rather than describing:

And Veryfi publishes a discount that makes that floor bite for longer, not less. The page offers: "Starter (Pay As You Go Plan): Commit to 12 months and get 1 cent discount for every document." (read 5 September 2026). A receipt then costs $0.07 rather than $0.08 — but the $500 minimum does not move, so the volume at which you finally stop overpaying rises from 6,250 documents a month to 7,143. The cheaper rate takes longer to consume the floor. That is arithmetic on the vendor's own two figures, and it is the opposite of how a discount is meant to read.

Docsumo's Starter behaved the same way while it was published — $499 bought a block of up to 5,000 pages, so a business processing 300 pages paid the same as one processing 4,900. In August 2026 Docsumo withdrew that published price entirely; its paid plans are now quote-gated, which does not change the block-pricing lesson — it just moves the vendor into the category where you cannot read the floor at all before a sales call.

Both are correctly priced for their intended customer: a business doing thousands of documents a month. They are simply not priced for anyone below that, and the pricing page does not say so out loud.

Free tiers are where most small users should start instead, and the 5 September re-read finally put numbers on them. *Docsumo's free plan is the most generous in this comparison by a wide margin — "Up to 1000 free pages", with "10 users licenses" — which this guide previously flagged as a figure the reader would have to go and confirm. Veryfi, whose $500 floor is the steepest in our ledger, also runs a free tier at "Up to 100 docs/month"*** that we had never recorded at all. Parseur's remains the one that will not hold still, for the reason set out above (all read 5 September 2026). If your volume is genuinely small, start free and stop reading — and start with the thousand pages, not the twenty.

The middle of the market is where it gets interesting

Mindee at $53/month billed annually is aimed at the volume most small businesses actually have: tens to low hundreds of documents a month. Over three years that is $1,908 — for a job that is, at bottom, read this receipt and give me the fields.

Parseur is back in this paragraph beside it. Its pricing page now prints $49 a month, or $468 a year billed annually ($39 a month), for 100 pages (read 24 September 2026). We took it out while that page showed no figure, because we are not going to carry a middle-of-market recommendation on a number we cannot re-read at the vendor; it prints one again, so it belongs here.

What the self-hosted option changes, and what it does not

Extracta is a self-hosted AI invoice extractor. $89.00, one time (product page, read 2026-08-20). It is one of the five bring-your-own-key tools in the AI Document Suite.

Try the Extracta demo ↗Live, on sample data, no sign-up.

It removes the meter and replaces it with a different one, and the guide will not pretend otherwise:

"Do I need my own API key?" — "Yes — an OpenAI or Anthropic API key, or a local Ollama instance, is required; Extracta doesn't include one."

"What does it cost to use?" — "Extracta itself is a one-time purchase; your ongoing cost is only your own LLM usage, which varies by provider and model and should be checked with your provider."

So the comparison is: a vendor's meter with the vendor's markup, versus your provider's meter at your provider's price. That is a genuine structural difference — you pay the model owner directly instead of a company that resells the model owner — but it is not "free," and the product page does not claim it is.

The exception is the one that matters most for sensitive documents:

"OpenAI, Anthropic, and Ollama" are all supported — and with Ollama running locally, there is no per-document charge and no third party sees the document at all.

For invoices, receipts and contracts, that is a data-path difference, not a price difference — and it is the only configuration in this whole comparison where the document never leaves your building.

And the key handling is specified rather than assured:

"It's stored in your own database and transmitted only in the outbound HTTPS request you trigger to your chosen provider — never logged, echoed, or passed through any Extracta infrastructure."

The limitation that will decide it for some readers

"What about PDFs?" — "Image formats (PNG, JPEG, WebP) work reliably with all three providers; Anthropic handles text-layer PDFs natively, while OpenAI and Ollama treat PDFs as best-effort."

If your documents arrive as PDFs and you want to run locally on Ollama, that combination is explicitly "best-effort." Read that as: convert to images first, or use Anthropic, or accept variable results. A vendor that publishes the weak combination rather than burying it has earned some trust, but the limitation is real and it is common — most supplier invoices arrive as PDFs.

The gap this category actually has

Look for self-hosted document extraction and what you find is OCR engines — Tesseract, EasyOCR, PaddleOCR. They are the recognition layer: they turn pixels into characters, and none of them extracts structured invoice data on its own. They have no invoice-domain awareness and no table-structure preservation, which pushes the line-item and field-mapping problems back onto your team.

So the self-hosted market here is a free engine or a metered API, with very little in between. A tool that supplies the domain layer — the prompt, the field mapping, the review queue — while calling a model you chose, sits in that gap. That is the honest description of what a one-time license buys here: not the intelligence, the plumbing around it.

Choosing, by volume

Under ~100 documents/month — *Docsumo's free plan at "Up to 1000 free pages", or Veryfi's at "Up to 100 docs/month"***. Pay nothing, and note that the vendor with the harshest paid floor gives away the second-largest free tier.

Small volume and you want hosted — Docsumo's free plan, whose thousand pages will outlast most small businesses' first year (paid pricing remains withdrawn: both paid tiers read "Talk to us" on 5 September 2026).

Tens to low hundreds, hosted, want zero infrastructure — Parseur or Mindee at $39–$53/month on annual billing. Note that is $1,400–$1,900 over three years.

Thousands per month — Veryfi or Docsumo (Docsumo now quote-gated, so budget a sales call). The floors stop being punitive at volume and the accuracy engineering is what you are buying.

Any volume, but the documents are sensitive, or you want to pay the model owner directly — self-hosted with your own key, and local Ollama if the document must not leave the building.

Nobody should self-host expecting a free lunch. You have swapped a subscription for a usage bill you now administer.

The honest costs

You own the accuracy problem. A hosted vendor tunes extraction across millions of documents. Your results depend on your model choice and your prompts. Human-in-the-loop review is not optional at any volume that matters.

Model pricing moves. Providers reprice; you re-check. The product page's own advice is to verify with your provider, and that is permanent advice, not one-time.

Backups, uptime, hosting. "PHP 8.1+ with pdo, curl, and fileinfo — present on virtually all cPanel/Plesk hosts." For a concrete number from a named provider: Hetzner's own pricing — machine-read on 10 August 2026 from the price API behind its web-hosting and cloud pricing pages — lists web-hosting plans from €1.60 to €16.72 a month and entry cloud servers from €3.99 a month (prices as returned for their German datacenters; confirm current pricing and VAT treatment at checkout). Any comparable host works — the point is that the "own" side of this comparison costs single-digit euros a month.

The short version

Work out your monthly volume first. It decides everything else.

Below twenty documents, use a free tier. In the hundreds, the hosted middle-market tools are $39–$44/month and perfectly reasonable. In the thousands, the $500 floors stop being a penalty and start being a fair price.

And if you are doing two hundred documents a month on a plan with a $500 minimum, you are paying thirty-one times the list rate for the privilege of being small. Check which side of the floor you are on before you renew anything.

Next steps

• Open the live Extracta demo — no sign-up, already signed in, resets itself
• Extracta product page and license options
• Converting bank statements — the question is who reads them — the same upload question, for the most sensitive document a business holds
• The AI Document Suite — invoices, receipts, spec sheets, freight documents and bank statements, five tools, your key
• The SaaS Price Observatory — Veryfi, Mindee, Docsumo, Parseur and every other tracked vendor, quoted from their own pages and dated

LLM running costs are not quoted here as a figure. They vary by provider and model and change without notice — check them with your provider, as the product page itself advises.