Mistral releases OCR 4.1 with block-level confidence scores

Mistral announced the public preview of OCR 4.1, the newest version of its optical character recognition service that powers the company's Document AI stack. The model, mistral-ocr-4-1, brings two major upgrades over previous versions: native paragraph-level bounding box extraction and structural block labels. This means the OCR doesn't just recognize text, but identifies logical units like headings, body paragraphs, tables, and lists, while providing block-level confidence scores. These features allow downstream systems to understand document structure, not just raw text, which is critical for automated document processing, compliance, and retrieval-augmented generation (RAG) pipelines.

OCR 4.1 is priced at €3.5 per 1,000 standard pages and €4.38 per 1,000 annotated pages. It supports a dedicated /v1/ocr endpoint with structured annotations, plus /v1/batch for asynchronous batching of large document volumes. The release also highlights compatibility with other Mistral models, including Z.ai GLM 5.2, Shieldstral 1.0, and Mistral Medium 3.5, enabling a full document-processing stack from OCR to classification and extraction. For developers, this unlocks cost-efficient, layout-aware document parsing at scale, making it easier to turn PDFs and scans into structured, machine-readable data.