Mistral OCR 4: Bounding Boxes, Pricing and API

Mistral OCR 4 is the document-reading model Mistral AI released on June 23, 2026, and it returns much more than text. Each page comes back as structured blocks, with a position, a type and a confidence score attached to each.

The headline figures, from the 85.20 OlmOCRBench score to the 72% win rate, are Mistral's own measurements. The API starts at $4 per 1,000 pages, and running it on your own infrastructure is limited to enterprise customers.

What Mistral OCR 4 is and what it returns

Mistral OCR 4 turns a document into text plus a structured map of the page: where each block sits, what type it is, and how confident the model is about it. OCR, short for optical character recognition, converts the characters in an image or PDF into text a program can use. Mistral says earlier generations concentrated on clean text and tables, while OCR 4 hands back a structured representation of the page.

What Mistral OCR 4 returns for one document

Input: a document such as PDF, DOC, PPT or OpenDocument

↓ Read by Mistral OCR 4

Text … Markdown-structured text

Bounding boxes … where each block sits on the page

Block types … titles, tables, equations, signatures and more

Confidence scores … inline, per page and per word

Bounding box OCR: where each block sits on the page

A bounding box is a rectangle that marks where a piece of text or a figure sits on the page. Mistral calls bounding boxes its most-requested capability. An application can use one to highlight the region a piece of text came from, and the OCR 4.0 model page describes the extraction as paragraph-level, so do not assume word-level precision.

Take a contract-review app answering a question about a termination clause. With a box, it can point at the clause on the original page instead of only quoting it. Check the API reference for the coordinate format before you plan any highlighting code.

Block types and per-word confidence scores

Each block is classified by type, and confidence scores are generated per page and per word. Mistral names titles, tables, equations and signatures as block types, and a selector on the announcement page also lists Title, Subtitle, Text, Image, Table and Thinking. Read together, the fields give a downstream system four things for every region: its text, its position, its role and how sure the model is.

Mistral names three uses for that structure. Classified blocks make cleaner retrieval units for RAG (retrieval-augmented generation, where an AI answers from your own documents), agents get typed pieces to act on, and ingestion pipelines get consistent typed output. Separately, Mistral says confidence scores let a data pipeline use human verifiers efficiently.

How Mistral OCR 4 differs from Mistral OCR 3

Mistral's announcement describes the change from earlier generations as a move from text-and-table conversion to a structured representation of the page. The new outputs, meaning boxes, block types and per-word confidence, are what the announcement presents as new.

Mistral compared OCR 4 against its own Mistral OCR 3 in its evaluations. If you are choosing between the two, run both on a sample of your own documents and compare the fields you actually need.

How Mistral OCR 4 scores on benchmarks

Mistral reports a top OlmOCRBench score, a 72% average win rate in a human evaluation and a lead on its own multilingual test, and it says the numbers are only directional. OlmOCRBench and OmniDocBench are public benchmarks that compare OCR output against reference annotations. Crawl Multilingual is Mistral's internal evaluation, so its result is a vendor figure by definition.

OlmOCRBench, OmniDocBench and the 72% win rate

OCR 4 scores 85.20 on OlmOCRBench, the top overall result among the models Mistral tested, and 93.07 on OmniDocBench. It also leads Mistral's internal Crawl Multilingual evaluation at 0.98. In the blind human evaluation, OCR 4 won more often than each competing OCR and document-AI product, averaging a 72% win rate.

Mistral OCR 4 results as reported by Mistral

The three bars use different scales: two are benchmark scores and one is a human-preference win rate, so bar length is not a ranking. Crawl Multilingual (0.98) is on another scale and stays in the text.

85.20OlmOCRBench score

93.07OmniDocBench score

72Avg. human-preference win rate (%)

0255075

The human evaluation used 600+ documents in 12+ languages from third-party vendors meant to reflect real industry use, and annotators compared each competitor's output with OCR 4's, one document at a time. Mistral says a human ranking on realistic documents avoids much of the annotation and formatting noise that affects automated scores.

Mistral also publishes two customer statements. An AI engineer at Rogo reports equivalent accuracy at roughly 8x lower cost and 17x lower latency than leading agentic document parsers on a chart-heavy financial QA dataset.

Anaqua reports roughly 4x faster processing per page than its incumbent provider. Both are statements Mistral chose to publish, not independent tests.

Why Mistral calls its own scores directional

Mistral says both OlmOCRBench and OmniDocBench have known limits in how they score some outputs, so one aggregate number can mislead in either direction. When it audited the mismatches behind its scores, most turned out to be artifacts of how the benchmarks compare output, not model errors. It lists five recurring causes:

• Errors in the reference annotations themselves.
• Different LaTeX that renders identically, counted as a mismatch.
• Equations split into fragments differently from the reference.
• Multi-column reading order, including words split across a column boundary.
• Header and footer handling: the benchmark does not expect them, so Mistral strips them before scoring, and the test then wrongly flags a string that also happens to be the page title.

These artifacts cluster in mathematical, scientific and multi-column documents, and Mistral says they more often penalize correct output than reward wrong output. It also notes that all competitor scores are its own internal reproductions. In practice, run a sample of your real files before you trust any leaderboard number.

Mistral OCR languages: 170 across 10 groups

OCR 4 supports 170 languages across 10 language groups. On the internal Crawl Multilingual test, Mistral says it leads in all eight language groups it reports, from English and the European and Middle Eastern groups through Chinese, East Asian and Southeast Asian languages to a specialized group. The specialized group includes Hindi, Japanese, Georgian and Tamil, among others.

Mistral says the margin is widest on specialized and low-resource languages, where many competing systems degrade sharply. Ten groups are supported, but the test reports results for only eight of them.

Mistral OCR pricing per 1,000 pages

Mistral OCR 4 costs $4 per 1,000 pages through the API, $2 with the Batch API, and $5 per 1,000 pages for Document AI. These are list prices from Mistral's announcement and its OCR 4.0 model page, as of September 30, 2026. The tiers differ in what they return and how the request is submitted.

Standard OCR, Batch API and Document AI prices

Standard OCR is $4 per 1,000 pages, and the Batch API's 50% discount brings it to $2 per 1,000 pages. The Batch API is the bulk route, aimed at high-volume ingestion where you want control over throughput and cost. Document AI is $5 per 1,000 pages, which the model page words as $5 per 1,000 annotated pages.

Mistral OCR 4 list price per 1,000 pages (USD)

The Batch API halves the standard rate. Document AI is the highest bar, and it is the tier that adds structured layers on top of the OCR result.

$024

$2Batch API

$4Standard OCR

$5Document AI

The OCR 4.1 card at the foot of the announcement page shows the same $4 and $5 list prices. Prices can change, so confirm them on Mistral's pricing page before you budget.

What the $5 Document AI tier adds

The $5 Document AI tier adds structured layers on top of the same OCR result, through the same endpoint with extra parameters. You can pass a JSON schema you define, and Mistral feeds the OCR output to mistral-small-2603 to generate content shaped to that schema. You can also annotate detected images with structured JSON, which triggers an additional vision-language model call per image, or add a custom prompt alongside a schema to steer how the whole document is interpreted or summarized.

Mistral Studio also offers Document AI as a no-code, application-level route to the same engine. It fits business users or pilot teams who want structured results without writing parsing logic.

How to use the Mistral OCR API

You call OCR 4 through a single endpoint, and the main choice is whether to add Document AI parameters to the request. The integration decision stays small, because the extraction result comes back either way. Mistral points new users to its Getting Started with OCR 4 Cookbook, which starts with a first extraction before moving on to boxes and block types.

One endpoint, plain OCR or Document AI parameters

Use OCR 4 as it is when you need the raw extracted content, and add the Document AI parameters to the same call when you need the output reshaped, annotated or driven by a custom instruction. Every request returns extracted text, bounding boxes, block types, confidence scores and Markdown-structured text, whatever you add. Plain extraction suits four cases: embedding OCR in an application, agent or pipeline; driving your own logic from the raw response; high-volume Batch API ingestion; and self-hosting for strict privacy or compliance needs.

Model name, endpoints and supported features

The model name on Mistral's OCR 4.0 model page is mistral-ocr-4-0, listed as version 4.0 and labelled GA and Premier. The page lists bounding box extraction, OCR and structured annotations on /v1/ocr, and batching on /v1/batch.

The announcement names PDF, DOC, PPT and OpenDocument as examples of the formats OCR 4 accepts. Every request returns the extracted content with its positions, block types and confidence scores.

Where the Mistral OCR API is available

Mistral OCR 4 and Document AI are available through the API in Mistral Studio, Amazon SageMaker and Microsoft Foundry, with Snowflake Parse Document listed as coming soon. As of September 30, 2026, the announcement page still lists Snowflake as coming soon, so do not plan around it until Mistral confirms. In the same announcement, Microsoft's Kimmi Grewal, VP of AI Ecosystem Partnerships, is quoted on Document AI with OCR 4 in Foundry.

Can you self-host Mistral OCR 4?

Yes for enterprise customers who arrange it with Mistral, through a self-managed deployment that runs in your own environment. Asking whether it is open source covers two questions: whether you can run it on your own infrastructure, and whether you can download it. Mistral's answer to the first is self-managed deployment for enterprise customers.

Single-container, self-managed deployment

OCR 4 is compact enough to run in a single container, which keeps document data inside your own environment. Mistral ties that to data residency, sovereignty and compliance, and says self-managed deployment is available to enterprise customers. Mistral asks anyone who wants to explore self-deployment to get in touch.

Mistral OCR 4 in Mistral Search Toolkit

Mistral Search Toolkit is an open-source, composable search framework in public preview, and OCR 4 serves as one of its ingestion components. Mistral announced the toolkit at the AI Now Summit.

OCR 4's structured output gives the toolkit citation-ready inputs for its ingestion, retrieval and evaluation workflow, aimed at RAG and enterprise search. For a separate Mistral retrieval piece, see Mistral Agentic Search: 3x on FinanceBench vs RAG, which covers the search side rather than document reading.

When to use Mistral OCR 4, and when not to

Use Mistral OCR 4 when you need documents turned into typed, located and scored content for RAG, agents or enterprise search, and skip it for anything that is a decision rather than a reading. Mistral itself makes this distinction in the announcement. As of September 30, 2026 its site also shows OCR 4.1, so check which version you are pricing.

Workloads Mistral names as a fit

Mistral recommends OCR 4 for complex multilingual document parsing, RAG, agent workflows, confidence-scored data pipelines and enterprise search. Here, agent workflows are tasks such as form filling, invoice processing and compliance checks, which Mistral says matter most in legal, financial services and healthcare. Early users report converting invoices into structured fields, digitizing company archives, pulling clean text out of technical and scientific reports, and running enterprise search on top.

Before you wire up an API, set a plain-text baseline: convert a sample PDF to Markdown and compare that output with what OCR 4 returns. A technical report with tables and equations makes a useful test file.

Out-of-scope uses Mistral lists

Mistral says OCR 4 is a document-understanding model, not a decision-maker. Out of scope, in Mistral's list, are medical diagnosis, legal advice or judgment, high-stakes financial decisions and safety-critical systems. Real-time or latency-sensitive processing is also excluded, as are inputs that are not documents, such as raw audio and video.

Because the model returns confidence values, a low-confidence region is a signal to send that region to a person rather than accept it as read. Mistral's own recommended use of confidence scores is to make human verification efficient.

Mistral OCR 4 or Mistral OCR 4.1?

As of September 30, 2026, Mistral's site lists OCR 4.1 alongside OCR 4.0, at the same $4 and $5 per 1,000 pages. The announcement page carries a card labeled OCR 4.1, marked New and described by Mistral as the world's best document extraction and understanding model. Check Mistral's model list for the current version before you pin a model name in production.