Mistral drops Mistral OCR to turn complex PDFs into clean data

Anyone who has built a RAG pipeline knows the pain of feeding PDFs into an AI: tables break, multi-column layouts bleed into each other, and mathematical formulas turn into unreadable soup.

Mistral has rolled out Mistral OCR, a dedicated optical character recognition and document understanding model designed to turn messy static documents into clean, AI-ready data.

Lightweight, insanely fast, and built for complex layouts

Mistral OCR goes far beyond legacy OCR software. It ingests PDFs and images, then outputs ordered, interleaved text and images formatted directly in Markdown.

It handles intricate layouts natively, transcribing scientific formulas into LaTeX, parsing dense tables, and extracting embedded charts and figures rather than just ignoring them.

Speed and pricing are where it really catches the eye. Built on a lightweight architecture, Mistral OCR can churn through up to 2,000 pages per minute on a single node.

The API is priced at $1 per 1,000 pages on standard inference, with batch processing cutting the cost roughly in half to around 2,000 pages per dollar. The tool is already the default document reader inside Mistral's Le Chat assistant, and API access is live on la Plateforme.

Outperforming frontier models on document benchmarks

In benchmark evaluations on Mistral's internal text-only test set, Mistral OCR scored 94.89 overall. That puts it notably ahead of GPT-4o at 89.77, Azure OCR at 89.52, and Google Document AI at 83.42.

The gap is especially pronounced in hard categories: it posted 94.29 on math, 96.12 on tables, and 98.96 on scanned materials.

Linguistically, the model boasts a 99.02% fuzzy match generation score across global languages. In Chinese tests, it reached 97.11%, compared to Azure OCR at 91.40% and Google Document AI at 90.89%.

Another neat addition is a capability Mistral calls "Doc-as-prompt." Users can prompt the model directly with a document to extract targeted variables and structure them into JSON schemas for downstream agent execution. For teams handling classified or tightly regulated records, Mistral is also offering selective on-premises self-hosting.

Why it matters

For engineers and data teams, document parsing has long been the dirtiest bottleneck in generative AI development. Piecing together open-source parsers, heuristic layout tools, and traditional OCR frequently leads to brittle pipelines. A unified API that extracts text, formulas, and visual assets for a fraction of a cent per page is an immediate win for enterprise RAG builds.

Across the industry, this highlights a clear pivot from monolithic frontier models doing general document tasks toward specialized, efficient multimodal models. Passing hundreds of pages through giant models like Gemini or GPT-4o is simply too expensive and slow for production archives.

We think Mistral found a massive sweet spot here. By attacking the mundane but essential task of converting static PDFs into clean structured Markdown, they are monetizing the enterprise data on-ramp. Offering on-premise deployment alongside a dirt-cheap API puts serious competitive pressure on legacy cloud providers like Azure and Google Cloud.