On June 23, 2026, Mistral AI released Mistral OCR 4, exposed in the API as “mistral-ocr-latest.” The company describes it as a small, focused model — compact enough to run in a single container — but the output is the story: every request comes back as Markdown, with bounding-box coordinates for each element, block-type classification (titles, tables, equations, signatures), and confidence scores at both page and word level.
In 2026, those fields matter more than raw recognition. RAG pipelines and agentic workflows treat documents as their primary input, and whether a pipeline can chunk reliably, retrieve, fill forms, or audit its own extraction quality depends on exactly this metadata. Pricing lands at $4 per 1,000 pages, dropping to $2 on the batch API.
From Text Extraction to Structured Output
The spec sheet: 170 languages across 10 language groups, including low-resource languages such as Georgian, Bengali, Tamil, and Kannada; input formats covering PDF, DOC, PPT, and OpenDocument; output pinned to Markdown with coordinates and block classification attached. Mistral is equally explicit about what the model is not for: medical diagnosis, legal judgment, high-stakes finance, safety-critical systems, latency-sensitive deployments, and non-document inputs all sit outside the design envelope.
The same endpoint also has a Document AI mode, enabled with extra parameters. It accepts custom JSON schemas, annotates images through a vision-language model, and takes custom prompts, with structured output generated by mistral-small-2603. In practice that folds OCR and schema-driven extraction into a single API call instead of a two-stage pipeline with a post-processing tier bolted on.
The use cases Mistral names are telling: RAG ingestion, enterprise search, semantic chunking, and agentic workflows like form filling, invoice processing, and compliance checks. All of them live or die on layout awareness and confidence metadata, not on raw character accuracy.
Benchmarks and Human Preference
The evidence comes in three layers. A human-preference evaluation across 600+ documents in 12+ languages, run as blind rankings, gave OCR 4 an average 72% win rate over every competitor tested. On the public OlmOCRBench it scores 85.20, the top result among tested models. On Mistral’s internal multilingual crawl benchmark it scores 0.98, placing first in all eight language groups. The OmniDocBench score of 93.07 arrives with an unusual caveat attached: Mistral says the benchmark contains ground-truth errors, LaTeX-equivalence traps, and column-ordering mismatches that can penalize output that is actually correct — a good reminder to read vendor benchmark pages all the way to the footnote.
Customer numbers are more concrete. Aidan Donohue of Rogo reports “equivalent accuracy at roughly 8x lower cost and 17x lower latency” versus leading agentic parsers. Ivan Mihailov of Anaqua calls it “roughly 4x faster per page than our incumbent provider.” Microsoft’s Kimmi Grewal marks the arrival of Mistral Document AI with OCR 4 in Microsoft Foundry as a milestone in the partnership.
Pricing and Deployment Options
Billing splits three ways: $4 per 1,000 pages on the OCR API, a 50% discount to $2 on the Batch API, and $5 per 1,000 pages for Document AI. Access runs through the Mistral Studio console, Amazon SageMaker, and Microsoft Foundry, with Snowflake’s Parse Document marked “coming soon.” Enterprises with data-sovereignty requirements can arrange self-hosted deployment through sales — notable for regulated European buyers, and consistent with Mistral’s positioning as the sovereign option. A webinar is scheduled for July 7, 2026, and a Getting Started cookbook is already available.
What It Means for Document Pipelines
The competitive frame has shifted. OCR is no longer a recognition-engine contest; it is a document-extraction API contest. Bounding boxes let downstream systems rebuild layout, per-word confidence enables quality gates before anything reaches a vector store, and Document AI mode removes an entire post-processing tier for teams that just need structured fields out of invoices and forms.
The Hacker News thread — 501 points, 136 comments — concentrated on two things: the pressure that $4-per-1,000-pages pricing puts on incumbent parser vendors, and the credibility debate sparked by the OmniDocBench caveats. The reasonable engineering conclusion is to treat OCR 4 as a candidate, not a verdict. Run the same methodology Mistral used — a blind human evaluation on your own document set — before committing a production pipeline to it.
Sources
AI-assisted summary compiled from the sources above, reviewed by a human before publishing.