For years, pulling data out of PDFs, invoices, and contracts meant choosing between accuracy and control. Cloud-based OCR services offered the accuracy but required you to send your most sensitive documents off-premise. On-premise tools offered the control but lagged years behind on quality.

Mistral AI’s OCR 4, released June 23, 2026, changes that calculation.

OCR 4 is a document intelligence model that does more than convert images to text. It returns structured representations of entire documents — bounding boxes, block-type classification, and per-word confidence scores — in a format that AI agents and downstream workflows can actually use. And critically for regulated industries, it deploys as a single container on your own infrastructure.

What’s Actually New

Most document processing tools extract text. OCR 4 extracts meaning.

When you feed it an invoice, a contract, or a financial statement, it doesn’t hand back a wall of text and leave you to figure out what belongs where. It classifies every block — headers, body text, tables, figures, addresses — assigns confidence scores to each word, and returns clean structured output compatible with retrieval-augmented generation pipelines, AI agents, and enterprise search systems.

The model supports 170 languages across 10 language groups and handles PDF, DOC, PPT, and OpenDocument formats — covering the practical reality of what businesses actually receive from clients, vendors, and counterparties.

Independent annotators gave OCR 4 a 72% win rate over every other system tested, placing it ahead of competing tools on structured extraction quality.

The Self-Hosting Angle

The most significant aspect for enterprise buyers isn’t the accuracy numbers. It’s the deployment model.

Regulated industries — healthcare, legal, financial services, accounting — can’t casually route client documents through a third-party API. Data processing agreements, HIPAA, GDPR, and internal security policies turn cloud document extraction into a compliance conversation before it becomes a technology decision.

OCR 4 ships as a single container deployable on an organization’s own infrastructure. That means extraction happens inside your security perimeter, with your data governance policies applied, and with no documents touching Mistral’s servers.

For accounting firms, law practices, and financial advisors who want to automate document processing but cannot send client data to external providers, this is the architecture that makes automation actually viable.

Pricing That Makes Piloting Easy

At $4 per 1,000 pages — dropping to $2 with the Batch-API discount — the economics are easy to evaluate. A firm processing 10,000 documents a month (invoices, contracts, financial statements) is looking at roughly $200 to $400 per month. For the time that saves on manual data entry alone, most organizations will find payback measured in days, not months.

What This Means for Business

Document processing is one of those unglamorous bottlenecks that costs real money across every professional services firm. Accounts payable teams manually re-keying invoice data. Legal teams manually reviewing contract terms line by line. Financial advisors extracting numbers from client statements before they can analyze anything.

Every one of those workflows is an automation candidate. The missing piece has typically been a document extraction layer accurate enough, fast enough, and safe enough to trust with sensitive client data.

OCR 4 addresses the accuracy and deployment gaps simultaneously. For businesses already building AI agent workflows, it removes one of the last friction points in the document intake layer. For businesses still evaluating where to start with AI automation, a self-hosted, per-page pricing model lowers the commitment to a level where running a pilot is straightforward.

Mistral is positioning OCR 4 as the document intelligence backbone for enterprise AI stacks — the piece that converts physical and digital documents into structured data that agents can act on. Given how much of business still runs on paper and PDFs, that market is larger than it might first appear.

If you’re building out AI agent workflows that handle documents — particularly in accounting, legal, or financial services — OCR 4 is worth evaluating. The combination of structured output quality, multilingual support, and self-hosted deployment options represents a meaningful step forward from what was available even six months ago.

Looking to add document automation to your business operations? Talk to the Omni team about what’s possible.