Mistral put OCR 4.1 in public preview with paragraph-level extraction

Why it matters

Paragraph-level extraction gives engineers finer document structure for search, agents and automated workflows. Documented page-level and word-level confidence data can help teams route uncertain results for human verification.

Mistral put OCR 4.1 into public preview on July 16, 2026. The model adds native paragraph-level bounding-box extraction, structural block labels and block-level confidence scores. Mistral's current OCR documentation describes page-level and word-level confidence output.

Arthur Mensch, Guillaume Lample and Timothee Lacroix founded Mistral in April 2023 after careers at Google DeepMind and Meta. Mensch became CEO, Lample took the chief science officer role and Lacroix became CTO. Their founding thesis centered on giving customers greater control over AI systems through accessible, customizable models. Document processing is a practical test of that pitch: companies need to know what a model extracted, where it found the information and how much trust to place in each result before an automated system acts on it.

The OCR 4.1 model page lists a price of EUR3.50 per 1,000 pages and EUR4.38 per 1,000 annotated pages. Those figures should not be treated as continuity with the previous release: Mistral's OCR 4 announcement priced API processing at $4 per 1,000 pages and Document AI at $5 per 1,000 pages. OCR 4.1 supports OCR, structured annotations, paragraph-level bounding-box extraction and batch processing through Mistral's Document AI stack.

The API documents three confidence layers

OCR 4, released on June 23, returned document blocks with bounding boxes and structural labels such as text, title, table, equation, code, header, footer and signature. The OCR 4.1 model page adds native paragraph-level bounding-box extraction and lists block-level confidence scores.

The current OCR documentation documents page and word as available confidence-score granularities. The OCR 4.1 model page separately lists block-level confidence scores, although the basic OCR guide does not explain that mode in detail. Page-level output provides aggregate confidence statistics, while word-level output adds scores for individual words. Separately, include_blocks=True adds a blocks array to each page with paragraph-level bounding boxes, structural labels and extracted content in reading order.

Mistral's changelog records the June 23, 2026 release of OCR 4 as mistral-ocr-4-0 and says mistral-ocr-latest points to it. The official documentation does not establish that a mistral-ocr-4 alias was redirected to OCR 4.1.

In its OCR 4 announcement, Mistral identified form filling, invoice processing, compliance checks, redaction, enterprise search and retrieval as intended uses. Mistral said confidence metadata could support human verification. Bounding boxes preserve the location of each block, while structural labels tell downstream software what kind of material it is handling. Those fields, combined with page-level, word-level and block-level confidence data, give engineers more granular inputs for review policies than an undifferentiated page of extracted text.

The benchmark case still belongs to OCR 4

Mistral published a detailed performance case for OCR 4. Mistral reported an 85.20 score on OlmOCRBench and said independent annotators preferred OCR 4 in 72% of comparisons across more than 600 documents. Mistral also cautioned that aggregate OCR benchmarks can misread correct equations, column ordering and block classifications as errors, and described the results as directional.

Those numbers apply to OCR 4. Mistral's published material for 4.1 describes its supported features without establishing a separate performance result. Customers evaluating the preview still need to test the model against their own scans, forms, tables, handwriting and mixed-layout documents. OCR 4.1's documented case rests on paragraph-level localization and structural metadata rather than a new accuracy benchmark.

The two releases also use different published pricing presentations and currencies. OCR 4.1's model page lists EUR3.50 per 1,000 pages and EUR4.38 per 1,000 annotated pages. The OCR 4 announcement listed $4 per 1,000 API pages, a 50% Batch API discount to $2 per 1,000 pages and $5 per 1,000 pages for Document AI.

Mistral wants the ingestion layer

Mensch and his co-founders are building beyond general-purpose language models. Mistral followed OCR 3 in December 2025 with OCR 4 in June 2026 and the OCR 4.1 public preview in July. Mistral is assembling focused models that can sit inside business and industrial systems.

Document ingestion is a useful entry point because enterprise agents depend on the quality of the material they receive. A model can reason well and still produce a bad result if a table was flattened incorrectly, a footer was confused with body text or a low-confidence figure entered the retrieval index without review. OCR 4.1 gives operators paragraph-level location, structural data and block-level confidence scores at that boundary, while the API documents page-level and word-level confidence metadata.

Mistral entered the OCR and document-intelligence market with the launch of OCR 4 on June 23, 2026. Rivals include Amazon Textract, which extracts text, handwriting, tables and forms, and Google Cloud Document AI, which combines OCR with layout analysis, classification and structured extraction. Those platforms benefit from their place inside large cloud portfolios. Mistral offers a focused document model connected to its own models, Studio tooling and enterprise AI stack.

OCR 4.1 is a narrow public preview by frontier-model standards. Its significance comes from where the founders placed the improvement: inside the document pipeline that feeds search, agents and automated decisions. Paragraph-level output gives developers finer structure to inspect, route and index. The API's documented page-level and word-level confidence data, alongside the model page's block-level confidence listing, provides additional inputs for evaluating uncertain extractions.