Why this matters now

Mistral OCR 3 is the newest release in Mistral’s Document AI stack, built to extract text and embedded images from complex documents with high fidelity and speed—now live in Studio and API. It reconstructs layout and tables (Markdown output enriched with HTML tables) and introduces highly competitive pricing for large-scale workloads.

Key points / benefits

• High-fidelity extraction of text + embedded images,a with structure retention (tables, layout) for easier downstream use.

• Fast and cost-efficient: listed at $2 per 1,000 pages (Batch API $1 per 1,000 pages effective), plus an annotated-pages option.

• Built for scale across invoices, forms, scans and mixed-quality documents; available via Mistral Studio and API.

How it works

Mistral OCR 3 upgrades the prior release (see Dec 2025 changelog) with improved layout understanding and HTML table reconstruction inside Markdown outputs—so systems can ingest both content and structure. It is also positioned as a smaller, faster model than typical enterprise OCR, enabling low latency and lower cost per page at volume.

Notable launch claims: Mistral highlights “state-of-the-art accuracy” and strong wins on forms, scans, complex tables and handwriting (as shared on official channels). Treat benchmark claims as vendor-reported until third-party evaluations land.

Practical steps / examples

• Bulk backfile conversion: run archived PDFs through OCR 3 using Batch API to minimise cost per page, then push structured Markdown/HTML into your ECM or data lake.

• Invoice & form capture: use table reconstruction to map line items directly into downstream schemas (AP, logistics, CRM) with fewer post-OCR regex rules.

• Knowledge workflows: extract interleaved text + images from research papers or contracts, then route to RAG pipelines with preserved section headings and tables.

• Human-in-the-loop QA: for regulated teams, sample annotated pages to spot-check accuracy before promoting pipelines to production.

Mistral OCR 3 is a high-fidelity document AI service that extracts text and embedded images while preserving structure (including HTML-reconstructed tables). It’s designed for speed and scale, with pricing from $2 per 1,000 pages and Batch API discounts for bulk processing.

FAQs

Q1: What makes Mistral OCR 3 unique?
It combines content + structure extraction (Markdown with HTML tables) in a smaller, faster model, enabling lower cost and latency at scale. mistral.ai

Q2: Does it handle multiple languages and messy layouts?
Mistral positions OCR 3 for diverse documents—forms, scans, complex tables and handwriting—and it’s part of a document-understanding stack used across multilingual content. Verify language coverage for your corpus during pilot. mistral.ai+1

Q3: How is it priced?
List price $2 / 1,000 pages; Batch API reduces effective price to $1 / 1,000 pages. An annotated-pages option is available in docs. Check your region and usage tier. mistral.ai+1

Q4: How do we access it?
Available now via Mistral Studio and API (model ID family: mistral-ocr-*, e.g., mistral-ocr-2512). See “Models” and “Changelog” for release details. docs.mistral.ai+1