Mistral OCR vs Azure Document Intelligence

Last updated July 2026.

Mistral OCR and Azure AI Document Intelligence solve different halves of the same problem. Mistral OCR 4 is a general document reader: it costs $4 per 1,000 pages, supports 170 languages, and returns each page as markdown with layout preserved. Azure's prebuilt invoice model costs about $10 per 1,000 pages, supports 27 languages for invoices, and returns typed invoice fields including line items without you writing a schema. If you need cheap, multilingual text at scale, Mistral wins on price. If you need vendor, invoice number, due date, and a line-item table straight out of the API, Azure does that job natively and Mistral needs its Annotations layer to match.

Both products moved fast in the last year, and most comparisons floating around are out of date. Mistral OCR 4 shipped on 23 June 2026 with a new model ID and a different price to the version people benchmarked in 2025. Azure's invoice model reached v4.0 GA and picked up optional key-value pairs. Here is where the two actually stand today for invoice and document work.

Mistral OCR vs Azure Document Intelligence at a glance

What is the difference between Mistral OCR and Azure Document Intelligence?

The difference is the unit of output. Mistral OCR returns the document as text: markdown that preserves headings, tables, and reading order, plus paragraph-level bounding boxes, typed-block classification for titles, tables, equations, and signatures, and inline confidence scores. It is a reader, and a very good one. Azure's prebuilt invoice model returns the document as data: named fields such as vendor name, invoice ID, billing address, due date, amount due, and a structured line-item array, delivered in the documentResults node of its JSON.

That distinction decides most of the architecture. With Azure you call one endpoint and get invoice fields. With Mistral you either write code that reads markdown and works out which number is the subtotal, or you use the Annotations feature and define the schema yourself, which Mistral bills as Document AI at $5 per 1,000 pages.

Which is cheaper, Mistral OCR or Azure Document Intelligence?

Mistral is cheaper for invoice extraction, roughly by half. Structured invoice output costs $5 per 1,000 pages through Mistral Document AI against about $10 per 1,000 pages for Azure's prebuilt invoice model. For plain text the ranking reverses: Azure's Read model runs about $1.50 per 1,000 pages while Mistral OCR is $4, although Mistral's Batch API discount of 50 percent brings it to $2 and closes most of that gap.

At real volumes both are inexpensive. Two thousand invoice pages a month costs about $10 on Mistral Document AI and about $20 on Azure. Neither number should decide an architecture on its own, and the engineering time around the API will dominate the bill either way. The detailed math for each sits on our Mistral OCR pricing and Azure Document Intelligence pricing breakdowns.

Does Mistral OCR extract invoice fields like vendor and total?

Not from the standard OCR endpoint. That returns markdown text with the layout intact, so the invoice total is present as characters on the page but is not labeled as the total. To get typed fields you use Mistral's Annotations feature, define a schema describing the fields you want, and pay the Document AI rate of $5 per 1,000 pages. It works well, and the schema gives you control Azure's fixed field set does not, but it is work you own. Azure's invoice model ships with that schema already defined and maintained by Microsoft.

Which is more accurate for invoices?

There is no clean head-to-head benchmark, so treat vendor claims carefully. Mistral reports a score of 85.20 on the public OlmOCRBench benchmark and human evaluator win rates averaging 72 percent against competing systems, but OlmOCRBench measures document text extraction rather than invoice field accuracy specifically. Azure publishes no equivalent single score for the invoice model.

What matters more in practice is failure mode. A general reader that transcribes a page perfectly can still hand you an ambiguous table where the tax line and the shipping line look identical, and your parsing code has to guess. A dedicated invoice model can misread a character but usually knows which field it is looking at. For accounts payable, the second failure mode is easier to catch, because a confidence score attached to a named field tells a reviewer exactly what to check. Test both on 50 of your own invoices, including your ugliest scanned ones, and count field-level errors rather than trusting either benchmark.

How many languages does each support?

Mistral OCR 4 supports 170 languages across 10 language groups. Azure's invoice model currently supports invoices in 27 languages, though Azure's general Read model covers considerably more. If you process supplier invoices from a wide international base, that gap is the single strongest argument for Mistral, and it is a large gap rather than a marginal one.

Can either one be self-hosted?

Both offer a route, and Mistral's is simpler. Mistral OCR 4 ships as a single container for self-hosted deployment, which suits organizations whose invoices cannot leave their own infrastructure. Microsoft also offers container deployment for Document Intelligence, subject to eligibility and separate licensing terms. If data residency is a hard requirement, both are worth a conversation, but Mistral's packaging is the more straightforward of the two.

What are the file and page limits?

Azure processes up to 2,000 pages per PDF or TIFF, with a 500 MB file size limit on the paid S0 tier and 4 MB on the free F0 tier. The free tier also reads only the first two pages of any PDF, which catches a lot of people out during evaluation: a multi-page invoice will look like it extracted badly when in fact it was truncated. Mistral accepts PDF, DOCX, PPTX, PNG, JPEG, and AVIF, giving it an edge if your documents arrive as Office files, since Azure's prebuilt models take PDFs and images only.

Which should you choose for accounts payable?

Choose Azure's prebuilt invoice model if you want invoice fields and line items out of the box, you are already on Azure, and your suppliers invoice in one of its 27 languages. You will write less code and inherit a field schema Microsoft maintains.

Choose Mistral OCR if you process many languages, want the lowest per-page cost, need documents in Office formats, or want a self-hosted container. Budget for the schema work, and remember that a homegrown extraction pipeline needs its own operational care: extraction quality drifts quietly when a big supplier changes their invoice template, so plan for the kind of monitoring that catches freshness and anomaly problems before a bad month of data reaches your ledger.

Choose neither if nobody on your team is going to write and maintain that code. Both are APIs that return JSON to an application you build. If the actual goal is invoices in a spreadsheet, a finished tool that already handles field extraction, invoice line item extraction, validation, review, and export is cheaper in practice even at a higher headline rate per page, because the developer time is the expensive part. If you want structured invoice JSON without building the surrounding pipeline, a flat-rate invoice data extraction API covers the same ground without per-page metering, and the Azure Document Intelligence alternative comparison walks through what that swap looks like in practice.

The short version

Mistral OCR 4 is the better document reader: cheaper, far more multilingual, broader format support, easy to self-host. Azure AI Document Intelligence is the better invoice extractor: it returns named fields and line items with no schema work. Mistral closes that gap with Annotations at $5 per 1,000 pages, still undercutting Azure's roughly $10, so a team willing to define and maintain a schema gets structured invoice data for about half the price. A team that would rather not should either pay Azure's premium or skip the build entirely.