Mistral AI says it has introduced Mistral OCR, an optical character recognition API aimed at document understanding rather than plain text extraction. In the company’s product announcement, OCR is described as a system that can handle images and PDFs while preserving the structure of media, text, tables and equations. The pitch is not just about reading words on a page. It is about turning dense documents into usable machine-readable inputs for downstream systems.
The company says the model is intended for organizations that work with complex files such as scientific papers, slides and technical PDFs. According to the announcement, Mistral OCR extracts content in an ordered, interleaved form that includes both text and images. That matters for retrieval and question-answering pipelines, where layout and visual elements can affect meaning. Mistral frames the product as especially useful when paired with retrieval-augmented generation systems that need multimodal documents rather than simple text blocks.
Mistral also says the model is available through its developer platform and already serves as the default document-understanding model across millions of users on Le Chat. The company says the API, called `mistral-ocr-latest`, is priced at 1,000 pages per dollar, with roughly twice the pages per dollar available through batch inference. The announcement says the service is available now on the company’s platform, with cloud, inference-partner and on-premises availability to follow.
Performance is a central part of the claim. Mistral says the model can process up to 2,000 pages per minute on a single node and can work across thousands of scripts, fonts and languages. It also says OCR can be used in self-hosted environments for organizations handling sensitive or classified data. That option suggests the company is pitching the product to enterprises that want to keep document workflows inside their own infrastructure rather than sending them to an external service.
The announcement further describes the system as document-aware in a way that allows users to use documents as prompts, extract specific information and format outputs as structured data such as JSON. Mistral says that makes the model useful for building agents and chaining extracted outputs into later function calls. It also lists use cases including research digitization, heritage preservation, customer support, and turning education, legal and engineering material into indexed, answer-ready content.
One detail in the source is that Mistral OCR is not a permanently frozen product. The company says it is no longer maintained and has already been replaced by newer OCR models. Even so, the announcement is still useful as a snapshot of how Mistral is positioning its document tools: fast, multilingual, layout-aware and suited to enterprise document pipelines rather than simple scanning. The larger message is that OCR is being treated less as a utility and more as a foundation layer for AI systems that need to reason over complex files.
