Mistral OCR Pricing and API Cost for Bank Statement Extraction

Mistral OCR 4 costs $4 per 1,000 pages through the standard API and $2 per 1,000 pages in batch mode. Mistral Document AI, the higher level product built on the same model, is $5 per 1,000 pages. Those rates make it one of the cheapest capable document OCR APIs you can buy in 2026. For bank statement work the price is almost never the thing that decides the project, and this article explains why.

How much does Mistral OCR cost?

Mistral prices OCR per 1,000 pages rather than per token, which is unusual for their catalog and makes budgeting simple. OCR 4 shipped on 23 June 2026 and is the current model. Batch mode applies the standard 50 percent discount, so if your workload can tolerate asynchronous processing you halve the bill with no change to the model.

Rates published by Mistral and verified on 29 August 2026 against the vendor announcement for OCR 4. Vendors change pricing, so confirm before you commit.

What does Mistral OCR actually return?

It returns the document as text with structure, not a spreadsheet. Alongside the extracted content you get bounding boxes, typed-block classification that labels titles, tables, equations and signatures, and inline confidence scores per block. Mistral states the model covers 170 languages and can run in a single container for fully self-hosted deployments, which matters if client financial documents are not allowed to leave your network.

Confidence scores are the genuinely useful part for financial work. Most OCR APIs hand you text and leave you guessing which parts to trust. A per-block score gives you something to threshold on, so a low-confidence amount can be routed to a human instead of silently landing in a ledger.

Is Mistral OCR accurate enough for bank statements?

Mistral publishes benchmark results for OCR 4: a score of 85.20 on OlmOCRBench, 93.07 on OmniDocBench, and human evaluators preferring it over competing models at win rates averaging 72 percent. Those are vendor self-published figures and should be read as such. To Mistral credit, their own announcement says benchmark scores contain artifacts and recommends evaluating on your own documents rather than relying on published numbers, which is the correct advice.

Take it. General document benchmarks are dominated by academic papers, forms and receipts. A bank statement is an unusual document: a long register printed across many pages, with repeated column headers, wrapped description lines, dates that omit the year, and summary tables that look structurally identical to the transaction table. A model can score extremely well on OmniDocBench and still hand you a statement whose totals do not foot.

Mistral OCR pricing compared to other document AI APIs

Here is the field at list price, per 1,000 pages, for the operation you would realistically use on a bank statement. Cheap text-only tiers are included because people reach for them and then discover they return no table structure.

Azure rates are East US from the public Azure retail prices API. Amazon Textract rates are US West Oregon, first 1 million pages, re-verified on 29 August 2026. All figures in US dollars at list price.

Mistral sits near the bottom of that table while still returning structure, which is a strong position. The two rows worth pausing on are the very cheap ones. Azure Read at $0.60 and Textract Detect Document Text at $1.50 look like bargains and are false economy for statements, because they return text lines with no table structure at all. You would be rebuilding columns from bounding box coordinates yourself.

How many pages is a bank statement, really?

People underestimate this badly, so the arithmetic is worth doing before you pick anything. A business checking account with a few hundred transactions a month commonly prints eight to twelve pages. That is roughly 100 to 140 pages for one account-year. Rebuilding three years of history for a loan file or an audit can pass a thousand pages for a single client.

Page counts are typical rather than guaranteed and vary by institution and transaction volume.

Read that table and the conclusion is uncomfortable for anyone hoping price would decide it. A bookkeeping practice processing 50 clients a month spends about $24 on OCR. The engineer configuring, testing and maintaining the pipeline around that API costs more in a single afternoon. At this end of the market the API rate is a rounding error and the build cost is the entire decision.

What the API price does not include

This is the part that turns a cheap line item into a project. Mistral gives you an excellent read of the page. Everything between that read and a workbook an accountant will sign off on is yours to build and keep building.

• Stitching pages into one register. A statement is a single continuous ledger printed across pages, with column headers repeating on each one that must be dropped rather than treated as transactions.
• Sign normalization. Some banks print separate debit and credit columns, some a single signed column, some a trailing CR marker. Each needs a rule, and the rules do not generalize between institutions.
• Dates with no year. Registers print 12/28 then 01/03. The year comes from the statement period, and a period crossing December means incrementing partway down the column.
• Wrapped descriptions. A long ACH memo spilling onto a second line arrives as a row with no amount. Merge it wrongly and you have invented a transaction.
• Summary blocks that are not transactions. Fee summaries, interest tables and daily balance grids are also tables on the page. A general model has no reason to treat them differently from the register.
• Footing the balance. Opening balance plus credits minus debits has to equal the printed closing balance. Without that arithmetic check, an extraction that drops one row in four hundred looks perfect and is wrong in a way that is expensive to find in an audit.
• Writing the output file. XLSX formatting, and QBO, OFX, QFX or QIF if the data is going into QuickBooks, Xero or Quicken. None of that exists in an OCR API.

None of these are hard in isolation. Together they are a small product with an owner, and every new bank layout arrives as an edge case without warning. This is the same list you would face building on Textract, Azure or an open-source parser, which is precisely why it is the real comparison rather than the per-page rate.

When Mistral OCR is the right choice

Pick it when an engineer owns the pipeline, the documents are varied rather than exclusively bank statements, and you want strong extraction at the lowest credible price per page. The single-container self-hosted deployment is a serious advantage for regulated teams, and it puts Mistral in the small group of options that can run inside a compliance boundary. If you are building broader enterprise document data extraction across contracts, invoices and statements at once, an OCR API with confidence scores is a sound foundation.

Pick a finished converter when the deliverable is a spreadsheet, the work is recurring, and the person doing it is a bookkeeper or an accountant rather than a developer. At 50 clients a month you would be spending an engineering budget to save $24 of API charges, which is a bad trade in any currency.

The honest summary

Mistral OCR 4 is well priced at $4 per 1,000 pages, $2 in batch, and it returns more useful structure than the cheap text-only tiers at Azure and AWS. It is a good OCR API. It is not a bank statement converter, and no OCR API is, because the difficult part of statement work is the accounting logic that sits after the text comes off the page.

If you want to compare the managed alternatives, the Azure Document Intelligence alternative covers the one service with a genuine US bank statement model, the Amazon Textract alternative covers AWS, and the Docling alternative covers the strongest open-source parser if you would rather self-host for free. For a wider per-page breakdown there is OCR API pricing for bank statements, and the Azure Document Intelligence pricing walkthrough goes meter by meter.

If you would rather skip the build entirely, the bank statement converter takes a PDF or scanned statement and returns Excel, CSV, QBO, OFX, QFX or QIF with the running balance already in its own column and the closing balance checked. It is free to try on a real file, and if you need it in your own software the bank statement converter API returns statement-shaped output directly.