Mistral OCR

What it is

Mistral OCR is a document-understanding API: it takes document pages as input and returns their contents in a structured, machine-readable form. It is not limited to plain text extraction — it comprehends the media, text, tables and equations that appear in a document^[1]. It also parses thousands of scripts, fonts and languages across all continents^[1], a property that matters for archives, contracts and paperwork in which several alphabets and layouts appear inside the same file.

Where it came from

Mistral AI released the API on 6 March 2025^[1]. The first model behind it carried the identifier mistral-ocr-2503, and that original model was retired on 31 December 2025^[3], with later versions taking its place.

Speed and cost

Mistral quoted throughput of up to 2,000 pages per minute on a single node, at roughly 1,000 pages per dollar; with batch inference the pages per dollar roughly double^[1]. For anyone digitising a large collection, these numbers are the practical part of the offer: throughput and price per page decide whether a big archive can be processed at all, long before quality questions are settled.

Benchmarks

Mistral reported an overall benchmark score of 94.89 for the model, against 89.77 for GPT-4o-2024-11-20 and 90.23 for Gemini-1.5-Flash-002^[1], and said it consistently outperformed other leading OCR models in rigorous benchmark tests^[1]. Aggregate scores of this kind deserve a careful reading. A single number hides which documents were in the test set — clean printed pages, phone photographs, handwriting, dense tables — and OCR quality usually varies more between document types than between models on a mixed benchmark.

Deployment and access

For organisations with strict privacy requirements, Mistral offered self-hosting selectively^[1]. That option speaks to a common obstacle in document work: the papers that most need transcription are often the ones — medical records, legal files, internal reports — that cannot be sent to a third-party service.

Later versions

The line continued with OCR 2 on 22 May 2025, OCR 3 on 18 December 2025, OCR 4 on 23 June 2026 and OCR 4.1 on 16 July 2026^[2]. Mistral's OCR page now states that the model is no longer maintained and has been replaced by newer OCR models^[1].

Where it sits

An OCR layer sits near the front of many document pipelines. Search indexes, retrieval systems and any analysis over scanned records begin by turning page images into text, so mistakes made at this step travel downstream: a misread table column or a dropped equation becomes a wrong answer later, with no obvious trace back to the conversion. That is why vendors compete on throughput, on how many scripts and fonts are covered, and on whether tables and equations survive the conversion rather than dissolving into loose characters^[1].

Mistral OCR was Mistral AI's entry into that layer as a standalone, priced API rather than a feature buried inside a chat product. Its measured claims — pages per minute, pages per dollar, a benchmark score against named competitors^[1] — are the terms on which document services are now compared, and its short life as a named model reflects how quickly that comparison is rerun. Within a little over a year the original endpoint went from launch to retirement^[3], replaced by numbered successors^[2].