Mistral's OCR 4.1 gives you the fields to route a document to a human. The price list is a different problem.

Mistral put OCR 4.1 into public preview on July 16, 2026, adding native paragraph-level bounding-box extraction, structural block labels and block-level confidence scores ^[1]. Those three fields are what a document pipeline needs to decide which pages a person has to look at, and they arrived alongside a price list quoted in euros that does not line up with the dollar figures published for OCR 4 ^[6][7][9].

Start with the useful part. OCR 4 already returned document blocks with bounding boxes and structural labels including text, title, table, equation, code, header, footer and signature ^[5]. Setting `include_blocks=True` adds a `blocks` array to each page carrying paragraph-level boxes, those labels and the extracted content in reading order ^[3]. That is the shape of a routing rule: a low-confidence block tagged `signature` on an invoice is a different escalation than a low-confidence `footer`, and with a bounding box you can hand a reviewer the crop instead of the whole page. Mistral's stated uses for OCR 4 were form filling, invoice processing, compliance checks, redaction, enterprise search and retrieval, and the company said confidence metadata could support human verification ^[14].

The gap is in the documentation. Mistral's current OCR documentation describes page-level and word-level confidence as the available granularities, where page-level gives aggregate statistics and word-level scores individual words ^[2]. Block-level confidence is listed on the OCR 4.1 model page, but the basic OCR guide does not explain that mode in detail ^[4]. If you are writing a threshold policy against block scores, you are writing it against a model-page line item, not a documented contract. Worth pinning before it becomes load-bearing in production.

Pricing is the second unresolved item. The OCR 4.1 model page lists EUR3.50 per 1,000 pages and EUR4.38 per 1,000 annotated pages ^[6]. The OCR 4 announcement listed $4 per 1,000 API pages, a 50% Batch API discount to $2 per 1,000, and $5 per 1,000 for Document AI ^[7]. The premium for annotation is roughly consistent across the two presentations at about 25% ^[8], but the currency change means these are not comparable numbers, and the source material itself warns against reading them as continuity ^[9]. OCR 4.1 supports batch processing through the Document AI stack ^[15], yet no euro batch rate is reported for it ^[19]. Anyone modelling cost on batch throughput has an open question.

The accuracy case also has not moved. Mistral reported an 85.20 score on OlmOCRBench for OCR 4 and said independent annotators preferred it in 72% of comparisons across more than 600 documents ^[10], while cautioning that aggregate OCR benchmarks can misread correct equations, column ordering and block classifications as errors and calling the results directional ^[11]. Mistral's published material for 4.1 describes features without establishing a separate performance result ^[12]. Version discipline is cleaner: the changelog records OCR 4 as `mistral-ocr-4-0` with `mistral-ocr-latest` pointing to it, and the documentation does not establish that a `mistral-ocr-4` alias was redirected to 4.1 ^[13].

Twenty-three days separate the OCR 4 release from the 4.1 preview ^[16], following OCR 3 in December 2025 ^[17]. The cadence is fast enough that the docs are trailing the model pages.