OCR API Pricing for Bank Statements: What Each Vendor Costs Per Page in 2026

Last updated July 2026.

For bank statements, the published OCR rates in 2026 run from $1.50 per 1,000 pages for plain text extraction up to $70 per 1,000 pages for AWS Analyze Lending, with the structured table and form parsers most people actually need sitting between $10 and $50 per 1,000 pages. The number that decides your bill is not the rate, it is your page count, and statements are far longer than the invoices these price lists were designed around.

Here is the comparison, followed by the arithmetic that changes the answer.

Published per-page pricing, July 2026

Figures taken from each vendor's own pricing page in July 2026. AWS rates are US West Oregon for the first million pages a month; Google rates are for the first million pages a month. Cloud rates vary by region and vendors change plans, so confirm before buying.

Azure is the outlier in that table in one useful way: it is the only one of the three big clouds with a prebuilt US bank statement model rather than a generic table detector, and it bills at the same $10 rate. The meter-by-meter detail, including the commitment tiers and the free-tier trap, is in the Azure Document Intelligence pricing breakdown, and the build-versus-buy side of it is on the Azure Document Intelligence alternative page.

The page-count math that changes the answer

Every price list above is quoted per page, and every mental model people bring to it comes from invoices. An invoice is one or two pages. A bank statement is not.

• A personal checking statement with modest activity: 2 to 4 pages a month.
• A business checking account with a few hundred transactions: 8 to 12 pages a month.
• One account-year of a business account: roughly 60 to 140 pages.
• A three year lookback for a loan file, a divorce matter, or an audit: often more than 1,000 pages for a single client.
• A bookkeeping firm catching up twenty client-years: 1,200 to 2,800 pages.

Run those through the table. Mindee's free 250 pages a month, generous for invoices, covers about two account-years of business statements. Textract's free Tables allowance of 100 pages a month covers one account-year, and only for three months. Google publishes no standing free allowance at all, so evaluation is paid from the first page.

At paid rates the picture flips: 2,800 pages through Textract Tables is about $42, and through Google Form Parser about $84. Those are small numbers. Which is exactly the point, and the reason the API line item is almost never what determines the cost of the project.

What the price list does not include

Every JSON-returning option in the table shares the same gap: none of them produce a spreadsheet. Between the API response and a file an accountant will sign off on sits a specific amount of work.

1. Deciding which detected table is the transaction register, when a page also contains a summary block and a fee recap.
2. Stitching the register across page breaks in order, without turning repeated header rows into transactions.
3. Normalizing signs, because banks variously use separate debit and credit columns, a single signed column, or a trailing CR marker.
4. Resolving dates that print no year, using the statement period from the header, including periods that cross a year boundary.
5. Merging wrapped descriptions that spill onto a second line with no amount.
6. Verifying the balance foots against the closing figure printed on the last page.

That is the real cost line. A first working version is typically a sprint. Keeping it working across dozens of bank layouts, as those banks redesign statements, is an ongoing owner. When people say an OCR API was cheaper on paper and expensive in practice, this is invariably what they mean.

How to compare fairly

Put the options on the same footing before you look at any number.

Which option fits which buyer

Engineering teams with continuous volume. Textract or Google Document AI. The per-page rate is trivial at scale, documents stay in your own account, and you already have people who can own a parser. Compare them on the Amazon Textract alternative and Google Document AI alternative pages.

Product teams embedding extraction. Mindee, or an API with statement-shaped output. Published per-page rates make unit economics predictable, which matters when extraction is part of what you sell.

Enterprises automating mixed document intake. Affinda or Rossum. You are buying a document operation, not a converter, and the annual figure is small against the headcount it replaces. The Affinda alternative and Rossum alternative comparisons lay out what each includes.

Bookkeepers, accountants, finance teams, and lenders doing the work by hand. A converter, because the deliverable is a file and the parser is already written. Uploading a PDF to a bank statement converter returns the XLSX with the running balance intact, and the same conversion exports QBO or OFX for the ledger. For anything that is not a statement, a general PDF to Excel conversion handles the simpler cases.

The test that settles it

Whatever you shortlist, run your worst statement through it. A scanned one, from a smaller bank, with a long ACH description that wraps. Then check three things: does the row count match the PDF, does the closing balance foot, and are debits and credits signed correctly. A tool that passes that test at $50 a month is cheaper than one that fails it for free, because the hour spent hunting a missing row costs more than either. Once the data is clean, the next step is usually to categorize the transactions for reporting.

Two options worth adding to any shortlist priced this way. Mistral OCR pricing is among the lowest of any API that still returns table structure, at $4 per 1,000 pages and $2 in batch mode, and at the other end of the model the Docling alternative is MIT licensed and free to self-host, where the cost moves entirely into engineering time and compute.

Two more worth pricing before you decide. The Reducto alternative publishes per-product rates that run from $15 per 1,000 pages for standard parsing to $40 per 1,000 for Deep Extract, with a 20 percent discount on batch jobs. The Unstructured.io alternative charges $0.015 per page after a 10,000 page free allowance, and its open-source library carries no per-page fee at all. DeepSeek prices in tokens rather than pages: its API runs from $0.22 per million input tokens off-peak, though its OCR model is not served there at all, which the DeepSeek OCR alternative explains. The best open source OCR for bank statements comparison covers the free self-hosted options alongside these paid ones.

One more worth naming, because its absence from any price list is itself the answer: MinerU operates a hosted parsing API but publishes no per-page rate. Its documentation describes a daily allowance of 1,000 pages at highest priority, with volume beyond that processed at lower priority rather than refused, and a lightweight Agent endpoint capped at 10 MB and 20 pages per document. If you need a committed rate you have to ask them, and the output is Markdown rather than a spreadsheet in any case, as covered in the MinerU alternative writeup.