Mistral's updated document-OCR model — paragraph-level bounding boxes, structural block labels, and block-level confidence scores.
Last verified · 2026-08-13 · by Moe Ameen
Mistral OCR 4.1 is an updated release of Mistral's optical character recognition and document-understanding model, from the French AI lab Mistral. It surfaced on July 16, 2026 as a public preview, served as `mistral-ocr-4-1` on Mistral's Premier tier, and Mistral describes it as the latest OCR service powering its Document AI stack. Like the rest of the line, it reads a document and returns a structured, machine-readable version of it rather than a flat wall of characters.
The refinements are in the structure. Mistral highlights native paragraph-level bounding box extraction, structural block labels, and block-level confidence scores — so each block of the page is localized, typed, and scored for reliability, with the extracted text formatted as clean markdown. It exposes OCR, bounding-box extraction, and structured annotations through the `/v1/ocr` endpoint and high-volume batching through `/v1/batch`. This is an iteration on Mistral OCR 4 (June 23, 2026); Mistral has not published a granular head-to-head against OCR 4, so treat 4.1 as a refinement of the same document-intelligence approach rather than a leap with new benchmark claims.
On pricing, Mistral's documentation lists OCR 4.1 at €3.5 per 1,000 pages for standard OCR and €4.38 per 1,000 annotated pages when you request the structured annotations. It continues the OCR 4 line, which covers 170 languages across 10 language groups and can be self-hosted on a single container for data residency. Because 4.1 is a preview, pricing and availability can shift — verify against Mistral's docs before you build on it.
One thing stays true across the whole OCR line: this is an extraction model, not a content generator. It reads documents and hands back structured text and layout. It does not write a social post, draft a blog, or make a video or an image. Turning what it pulls out into published content is a separate job.
The upgrade in OCR 4.1 that matters for a creator is fidelity of structure. Paragraph-level bounding boxes and structural block labels mean the markdown you get back keeps the document's tables, figures, and section breaks intact instead of mashing a report into one paragraph. That clean structure is exactly what makes the downstream generation accurate — hand OCR 4.1's markdown to Kompozy as a source and the engine's data-heavy formats track the real document. A quarterly report becomes an Infographic Photo poster whose numbers actually match the source table; a research paper becomes a Carousel that walks its findings in order; a whitepaper becomes a blog article and an email newsletter, all written in your voice through the Persona Brief.
Where a rough copy-paste would leave gpt-image guessing at a mangled figure, a structured OCR 4.1 extraction gives Kompozy a source it can trust — so the Infographic Photo and brand-exact HyperFrames carousel reflect the document rather than drifting around it. From there Kompozy schedules and publishes the whole set across the nine connected platforms from one queue, turning a dense document nobody on social would read into the week's posts.
Mistral OCR 4.1 is an updated OCR and document-understanding model from Mistral, surfaced July 16, 2026 as a public preview (mistral-ocr-4-1). It returns markdown-structured text with native paragraph-level bounding boxes, structural block labels, and block-level confidence scores, and powers Mistral's Document AI stack.
OCR 4.1 is a preview iteration on OCR 4 (June 23, 2026) in the same document-intelligence line, emphasizing paragraph-level bounding box extraction, structural block labels, and block-level confidence scores. Mistral has not published a granular head-to-head, so treat it as a refinement rather than a benchmarked leap.
Mistral's documentation lists OCR 4.1 at €3.5 per 1,000 pages for standard OCR and €4.38 per 1,000 annotated pages for structured annotations. As a preview model, pricing can change — confirm current rates in Mistral's docs.
No. OCR 4.1 extracts structured text and layout from documents; it does not generate content or media. To turn the extracted text into posts, carousels, infographics, blogs, newsletters, or video, pass it to a content engine like Kompozy, which generates each format in your voice and publishes across platforms.
It continues the OCR 4 line, which covers 170 languages across 10 language groups, with reported gains on rare and low-resource scripts. It can also be self-hosted on a single container for data residency.