Mistral Releases OCR API at $1 per 1,000 Pages, Claims 94.89% Accuracy on Document Benchmarks

Mistral AI has released an OCR (Optical Character Recognition) API priced at $1 per 1,000 pages, with approximately double the pages per dollar available through batch inference. The model is now live on Mistral's la Plateforme developer suite and deployed as the default document understanding model for Le Chat.

Performance Claims

According to Mistral, their OCR model achieves 94.89% overall accuracy on internal benchmarks, outperforming competing models:

• Mistral OCR 2503: 94.89% overall
• GPT-4o (2024-11-20): 89.77%
• Azure OCR: 89.52%
• Gemini 1.5 Pro 002: 89.92%
• Gemini 2.0 Flash 001: 88.69%
• Google Document AI: 83.42%

The company reports particularly strong performance on mathematical content (94.29%), tables (96.12%), and scanned documents (98.96%). However, these benchmarks are based on Mistral's internal "text-only" test set containing publication papers and web PDFs.

Technical Capabilities

Mistral OCR accepts images and PDFs as input and extracts content as ordered, interleaved text and images in markdown format. The model handles complex document elements including mathematical expressions in LaTeX, tables, multilingual text, and embedded imagery.

The system processes up to 2,000 pages per minute on a single node, making it what Mistral claims is the fastest in its category. The model supports thousands of scripts, fonts, and languages, with particularly high accuracy scores on European languages: German (99.51%), Spanish (99.54%), Italian (99.42%).

Multilingual Performance

Mistral reports 99.02% fuzzy match accuracy in multilingual generation, compared to Azure OCR (97.31%) and Gemini 2.0 Flash (96.53%). Language-specific scores include Russian (99.09%), French (99.20%), Hindi (97.55%), and Chinese (97.11%).

Features and Deployment

The API supports document-as-prompt functionality, allowing users to extract specific information and format outputs as structured JSON for downstream function calls and agent workflows. Unlike some competing models, Mistral OCR extracts embedded images alongside text.

Mistral will offer selective self-hosting options for organizations with data privacy requirements, keeping sensitive documents within customer infrastructure. The model will be available through cloud and inference partners, with on-premises deployment coming soon.

What This Means

Mistral's entry into OCR represents a significant price point at $1 per 1,000 pages, undercutting many existing document processing services. The claimed accuracy improvements—particularly the 5-6 percentage point lead over GPT-4o and Gemini on Mistral's internal benchmarks—would be substantial if validated on independent test sets. The extraction of embedded images alongside text differentiates it from pure text OCR systems, making it more suitable for RAG pipelines processing complex documents like scientific papers and technical manuals. The multilingual capabilities and speed (2,000 pages/minute) position it for high-volume enterprise document processing, though real-world performance will depend on document complexity and infrastructure.