Summary: Mistral AI released Mistral OCR 4 on June 23, 2026 as a document model returning Markdown with bounding boxes, block labels, and confidence scores. It is priced at $4 per 1,000 pages, $2 through batch processing, and $5 for Document AI, targeting RAG and agentic extraction. Vendor and customer benchmarks support the candidate but still call for testing on each team’s own documents.
On June 23, 2026, Mistral AI put out Mistral OCR 4, listed in the API as “mistral-ocr-latest.” The company pitches it as a small, focused model — light enough to live inside a single container — but what sets it apart is the output: every request comes back structured as Markdown, each element carries its bounding-box coordinates plus a block-type label (titles, tables, equations, signatures among them), and confidence scores are given at both page and word level.
In 2026, those fields count for more than raw recognition. Documents serve as the primary input to RAG pipelines and agentic workflows, and whether a pipeline can chunk reliably, retrieve, fill forms, or audit its own extraction quality hinges on exactly this metadata. On price, the model sits at $4 per 1,000 pages, and the batch API halves that to $2.
Output Shifts from Text to Structure
The spec sheet reads as follows: 170 languages arranged into 10 language groups, low-resource languages such as Georgian, Bengali, Tamil, and Kannada included; PDF, DOC, PPT, and OpenDocument accepted as inputs; Markdown as the fixed output format, coordinates and block classification coming attached. Mistral is just as clear about the boundary on the other side — medical diagnosis, legal judgment, high-stakes finance, safety-critical systems, latency-sensitive deployments, and non-document inputs all fall outside the design envelope.
The same endpoint also carries a Document AI mode, switched on through extra parameters. It takes custom JSON schemas, leans on a vision-language model to annotate images, and accepts custom prompts, with mistral-small-2603 generating the structured output. A single API call ends up handling both OCR and schema-driven extraction, eliminating the two-stage pipeline with its bolted-on post-processing tier.
The use cases Mistral lists say something on their own: RAG ingestion, enterprise search, semantic chunking, and agentic work like form filling, invoice processing, and compliance checks. Every one of them stands or falls on layout awareness and confidence metadata rather than on raw character accuracy.
Scores, Rankings, and Human Preference
Mistral’s evidence comes stacked in three layers. A blind human-preference ranking spanning 600+ documents in 12+ languages put OCR 4 at an average 72% win rate against every competitor tested. The public OlmOCRBench returns 85.20, the best result among the models tested. Mistral’s internal multilingual crawl benchmark gives it 0.98 and first place in each of the eight language groups. Then comes OmniDocBench at 93.07, carrying the kind of caveat vendors rarely write: Mistral points to ground-truth errors, LaTeX-equivalence traps, and column-ordering mismatches in the benchmark that can penalize output which is actually correct — vendor benchmark pages deserve to be read all the way down to the footnote, and this is the textbook case.
Customers supply harder numbers. Aidan Donohue of Rogo reports “equivalent accuracy at roughly 8x lower cost and 17x lower latency” against leading agentic parsers. Ivan Mihailov of Anaqua measures it at “roughly 4x faster per page than our incumbent provider.” At Microsoft, Kimmi Grewal calls the arrival of Mistral Document AI with OCR 4 in Microsoft Foundry a milestone in the partnership.
What It Costs, Where It Runs
Three price lines apply: $4 per 1,000 pages on the OCR API; the Batch API takes 50% off for $2; and Document AI runs $5 per 1,000 pages. Availability spans the Mistral Studio console, Amazon SageMaker, and Microsoft Foundry, while Snowflake’s Parse Document carries a “coming soon” label. Enterprises facing data-sovereignty requirements can arrange self-hosted deployment through sales — a meaningful detail for regulated European buyers, and in step with Mistral’s positioning as the sovereign option. A webinar sits on the calendar for July 7, 2026, and a Getting Started cookbook is already out.
Consequences for Document Pipelines
The contest has been redefined. Recognition engines no longer decide who wins at OCR; document-extraction APIs do. Bounding boxes let downstream systems rebuild layout, word-level confidence makes quality gates possible before anything enters a vector store, and Document AI mode spares teams that only need structured fields out of invoices and forms the trouble of running a post-processing tier.
On Hacker News, the thread collected 501 points and 136 comments and kept returning to two subjects: the strain $4-per-1,000-pages pricing places on incumbent parser vendors, and the credibility fight set off by the OmniDocBench caveats. For teams assembling a RAG pipeline, the sensible conclusion is to hold OCR 4 as a candidate, not as a settled answer — first rerun Mistral’s own method, a blind human evaluation over your own document collection, and only then decide whether a production pipeline should rest on it.