Document AI Pricing per 1,000 Pages: Textract, Azure, Google and Mistral OCR

Pulling named fields out of 10,000 form pages costs $500 on AWS Textract, $300 on Google's Form Parser and $50 on Mistral Document AI, at list price. Plain text OCR runs the other way: the three hyperscalers all charge $1.50 per 1,000 pages and Mistral charges $4. Document AI pricing is not one market. It is two, and the cheapest vendor flips between them.

Document AI pricing: four rate cards in one unit

Each vendor prices in a different unit, which is the first reason buyers get this wrong. AWS quotes per page. Azure and Google quote per 1,000 pages. Mistral quotes per 1,000 pages but splits plain OCR from structured output, which it calls Document AI.

Everything below is converted to US dollars per 1,000 pages at the first public volume tier. AWS figures are US West (Oregon), taken from the worked examples on the Textract pricing page, because its price grid does not render without a browser. Azure figures come from Microsoft's public retail price API for East US, because the pricing page itself shows placeholders. Google figures are its list prices. Mistral figures come from its own product announcements.

That sourcing detail matters more than it looks. Two of four vendors make it hard to read a price without running a calculator. You should treat that friction as information about how the vendor expects you to buy, which is through a sales conversation.

What each vendor actually sells. AWS splits Textract into APIs: DetectDocumentText for raw text, AnalyzeDocument for tables, forms, queries and signatures, and specialised APIs for expenses, identity documents and lending files. Feature charges stack, so a page analysed for forms and tables pays both.

Azure Document Intelligence sells Read, a family of prebuilt models including Layout and Invoice at one shared price, and custom models you train. Google Document AI sells Enterprise Document OCR, a Layout Parser, a Form Parser, specialised parsers and a generative Custom Extractor. Mistral sells one OCR model plus a structured-output layer on top.

The one rate card that is not a rate card. Mistral's figures do not sit on its main pricing page, which says only that OCR is billed per 1,000 pages. They sit in the OCR 4 launch post of 23 June 2026, which labels the model OCR 4 in the text and OCR 4.1 on its pricing card. If you are building a procurement file, screenshot that page. Launch posts get edited less carefully than pricing pages, and this one is your only primary source.

Plain OCR pricing is a commodity at $1.50

For raw text extraction, the three hyperscalers have converged on an identical list price. Textract's DetectDocumentText costs $0.0015 per page for the first million pages. Azure Read costs $1.50 per 1,000 pages. Google Enterprise Document OCR also lists at $1.50 per 1,000 pages up to 5 million pages a month.

Mistral OCR 4 lists at $4 per 1,000 pages, falling to $2 through its batch API. That makes it more expensive than every hyperscaler for raw text at standard rates, and still 33% dearer at batch rates. Its predecessor, OCR 3, launched in December 2025 at $2 per 1,000 pages and $1 in batch, so Mistral doubled its list price between versions.

My position is that nobody should pick an OCR vendor on raw text price. At $15 for 10,000 pages, the bill is smaller than the engineering hour you spend comparing it. The real question is what the output looks like and how much cleanup it needs downstream.

What the extra $2.50 buys on Mistral. Mistral returns markdown with tables and structure preserved, and OCR 4 added bounding boxes, typed block classification and inline confidence scores, according to its launch post. Hyperscaler Read APIs return lines, words and coordinates. If your next step is feeding text into a language model for retrieval, markdown may save you a parsing stage. If your next step is a rules engine that needs coordinates, it saves nothing.

What 10,000 pages costs by document type

This is the table the rate cards do not give you. I priced 10,000 pages of three common enterprise document types at each vendor's first public tier, using the cheapest product that does the job as described. Where two products plausibly fit, both are shown.

Multiply each figure by 10 for the cost of 10,000 pages. AWS is US West (Oregon), Azure is East US retail, Google and Mistral are global list prices. Azure Query Fields has two meters in the retail API, at $10 and $200 per 1,000 pages; the table uses the newer $10 meter dated November 2024. Google's Invoice Parser figure comes from its pricing page as returned in search, because the page did not load for direct reading.

Invoices: four vendors at one price

Invoice extraction is the most deployed of these workloads and the most fiercely competed, and the price shows it. Textract's AnalyzeExpense costs $0.01 per page, the same $10 per 1,000 pages that Azure charges for any prebuilt model and Google charges for its Invoice Parser. For 10,000 invoices you pay $100 on any of the three.

Mistral Document AI does the same job at $50 for 10,000 pages, but with a structural difference. The hyperscalers ship a fixed invoice schema trained on invoices. Mistral asks you to define the schema yourself, and a general model fills it. You trade a pretrained invoice model for flexibility, and you take on the job of proving it extracts your suppliers' invoices correctly.

Contracts: mostly a layout problem

Contracts are long, text-heavy and mostly free of key-value pairs. The useful output is clean text with headings and tables intact. Textract includes Layout free when you buy Tables, so the effective price is $15 per 1,000 pages. Azure Layout and Google's Layout Parser both cost $10.

If you want specific clauses pulled out, such as the renewal date or the liability cap, the price roughly doubles. Textract Queries add $5 per 1,000 pages to Tables. Azure's Query Fields add $10. Google pushes you to the Custom Extractor at $30, its most expensive tier.

Forms: the widest spread in the market

Forms are where Textract is priced well above everyone. Forms alone cost $0.05 per page, and forms plus tables plus queries cost $0.07. On 10,000 pages that is $500 or $700. Google's Form Parser and Azure custom extraction both cost $300. Mistral Document AI costs $50, which makes Textract's full feature stack 14 times the Mistral price.

A blended 30,000-page month on each document AI API

Few teams process one document type. A realistic back-office workload mixes them, so I built one: 10,000 invoice pages, 10,000 contract pages needing layout and tables, and 10,000 form pages needing named fields. Each vendor gets its cheapest product that does each job, at list price, with no commitment discount.

The working is simple multiplication of the table above. AWS: $100 for invoices, $150 for contracts and $500 for forms, total $750. Azure: $100, $100 and $300, total $500. Google: $100, $100 and $300, total $500. Mistral: $50 for invoices through Document AI, $40 for contracts through OCR 4 and $50 for forms, total $140, or $120 with contracts sent through batch.

Why the totals mislead on their own. A $610 monthly difference between AWS and Mistral is real money for a startup and invisible for an enterprise. If one accounts payable clerk spends an extra few hours a month fixing extraction errors, that labour costs more than the entire API bill on any of the four vendors. The cost that matters is price per page plus the cost of the errors that page produces.

That is why the price table needs an accuracy column, and why the next section is the most important one in this post. The pattern is familiar from how vendors meter agent tasks: the visible meter is cheap to compare and the expensive part sits outside it.

Document AI accuracy claims, and how each was measured

Every vendor in this comparison tells you its product is accurate. Only one publishes numbers, and those numbers come from tests the vendor designed and ran. Here is what is actually on the record.

Mistral's numbers are the only ones, and that is a weakness

Mistral's OCR 4 post reports a 72% average win rate in blind human ranking across 600+ documents. It also reports 85.20 on OlmOCRBench and 93.07 on OmniDocBench, both public benchmarks. The same post says plainly that all competitor scores are internal reproductions and that both benchmarks have known scoring limitations. Credit to Mistral for writing that down. It still means the comparison was run by the party that benefits from it.

The March 2025 table is the only place I found Azure and Google scored by name. It put Mistral at 94.89, Azure OCR at 89.52 and Google Document AI at 83.42 on an internal set. That model is now deprecated, the hyperscaler products have shipped new versions since, and at least one document-parsing vendor reported at the time that it could not reproduce Mistral's figure on its own documents. I would not use any of those three numbers in a buying decision today.

Why the hyperscalers publish nothing

AWS, Microsoft and Google sell accuracy through confidence scores, not headline percentages. Each field comes back with a confidence value, and you set the threshold below which a human reviews it. That is a defensible design, because accuracy on your invoices is the only accuracy that matters to you.

It also means nobody can compare them without running their own test. Third-party blogs quote field-level accuracy figures for each hyperscaler, but the ones I found did not publish their sample, document mix or scoring rule. I have left them out on purpose. Benchmarks that disagree on method produce rankings that disagree on order, a problem examined in why AI leaderboards disagree with each other.

Volume tiers and commitment discounts change the order

List price is the entry point, not the price a high-volume buyer pays. Every vendor discounts, and the discounts are not symmetrical.

Above one million pages a month

Plain OCR drops sharply at volume on all three hyperscalers. Textract DetectDocumentText falls to $0.0006 per page after one million pages, Azure Read falls to $0.60 per 1,000, and Google Enterprise OCR falls to $0.60 per 1,000 above five million pages. At that tier, Mistral's $2 batch price is more than three times the hyperscaler rate.

Structured extraction discounts less. Textract Forms falls from $0.05 to $0.04 per page, a 20% cut. Google's Form Parser and Custom Extractor fall from $30 to $20 per 1,000 pages above one million, a 33% cut. Even at those rates, Google's Form Parser at volume costs four times Mistral Document AI at list.

Azure commitment tiers

Azure sells prepaid monthly commitments. The retail price API lists a 20,000-page prebuilt tier at $190 a month, and a one-million-page tier at $7,500, which works out to $9.50 and $7.50 per 1,000 pages. Custom extraction commitments run from $27 down to $18 per 1,000 pages. If your volume is steady, the commitment saves 5% to 25%. If it is seasonal, you pay for pages you never send.

Google offers a savings plan that cuts Form Parser and Custom Extractor to $27 per 1,000 pages on a one-year term and $24 on three years. AWS publishes no commitment tier for Textract on its pricing page. Mistral publishes none either. For both, the volume price is whatever your account team agrees.

Where this comparison is weakest

This post compares rate cards. A rate card is a vendor's opening offer, and four things about it make the comparison weaker than the clean table suggests.

Mistral's low price comes with a schema you must own

Mistral Document AI fills a JSON schema you write, using a general language model on top of OCR. The hyperscalers' prebuilt invoice and expense models were trained on those document types and return a fixed set of fields. A pretrained model can be wrong, but it is wrong in ways the vendor has seen before. A general model filling your schema can be wrong in ways nobody has tested. On the evidence available, I cannot tell you which fails more often on your documents, and neither can anyone else who has not tested them.

Prices are regional and dated. The AWS figures are Oregon, Azure is East US, and both vary by region. Azure's retail API also carries two conflicting Query Fields meters, at $10 and $200 per 1,000 pages, and I used the newer one. Mistral doubled its OCR list price between versions 3 and 4 in six months. Every figure here should be re-checked before it goes in a budget, and this post will be re-verified on a 90-day cycle.

Product churn is a cost the table cannot show

Google announced in February 2026 that legacy Document AI processors would be discontinued on 30 June 2026, and pointed customers to newer foundation-model versions. Mistral has deprecated its first OCR model within about 15 months of launch. Each migration means re-testing extraction accuracy on your own corpus. If you are on a pinned version, that cost lands whether or not the price changes, which is the same exposure covered in the real cost of switching model vendors.

The case for paying Textract's premium. If you already run on AWS, Textract keeps documents inside your account, your network controls and your existing data processing terms. That is worth something to a regulated buyer, and the $450 saved on 10,000 form pages will not fund a new vendor security review. Mistral now offers self-hosting for OCR 4, which narrows this gap, but self-hosting moves the cost from the invoice to your infrastructure team.

Which document AI API to pick for each workload

Here is what I would do, stated as a position rather than a ranking. Pick by workload, not by vendor, and accept that the answer may be two vendors.

Invoices on a hyperscaler you already use. Price is identical at $10 per 1,000 pages across all three, the models are pretrained on invoices, and confidence scores give you a ready-made review threshold. Test Mistral Document AI only if your invoice volume runs into the millions, where its $5 rate starts to matter.

Contracts through layout, not extraction. Azure Layout or Google's Layout Parser at $10 per 1,000 pages, or Mistral OCR 4 in batch at $2 if the output is going into a retrieval pipeline. Pay for clause extraction only on the clauses you will act on.

Forms are where you should run a bake-off. The price spread is 14 times, the accuracy evidence is absent, and the volume is usually high enough to justify a week of testing. Score each vendor on the same 200 pages of your own forms before you sign anything. The method in how to build an evaluation suite your team will maintain applies directly.

The inference cost under all of these is falling, as tracked in the inference provider rate card. Mistral's pricing shows that falling compute costs do not reach every product line at the same speed. The forms premium at the hyperscalers reflects what buyers have been willing to pay, not what extraction costs to run.

Frequently asked questions

Where to start this week

Pull last month's page counts from your current document pipeline and split them by type: invoices, contracts, forms, other. Multiply each by the rows in the table above. Most teams find that one document type is most of the bill, and that is the only type worth renegotiating or re-testing.

Then take 200 pages of that type, with fields already keyed by a human, and run them through two vendors. Count wrong fields per page, not a vendor's percentage. That single number, multiplied by what it costs a person to fix one field, will tell you more than any rate card in this post.

Related reading