Mistral OCR: what it is, what it costs, and how to use it without code

What is Mistral OCR?

Mistral OCR is an optical character recognition API from the French AI company Mistral AI. The first version launched on March 6, 2025, took images and PDFs as input and returned the text and images of each page in reading order[2]. Mistral said it read thousands of scripts, fonts and languages, and could process up to 2,000 pages a minute on a single node[2].

Mistral OCR 3 followed on December 17, 2025, under the model ID mistral-ocr-2512[1]. Mistral describes it as reading cursive and handwriting written over printed forms, rebuilding tables with merged cells and headers, and handling scans with skew, low resolution and background noise[1]. Its models page now lists OCR 4.1 as the newest model, with paragraph-level bounding boxes and block labels, and keeps OCR 3 for existing integrations[3].

Mistral OCR models at a glance

OCR 4.1

Latest model, adds paragraph boxes, block labels and block confidence scores[3]

OCR 3

Model ID mistral-ocr-2512, released December 17, 2025[1]

mistral-ocr-latest

The API alias that points to the newest OCR model[4]

Inputs

PNG, JPEG and other images; PDF, PPTX, DOCX and other documents[4]

Output

JSON with Markdown for each page, plus images, tables, headers and footers[4]

Price

$2 per 1,000 pages for OCR 3, $1 with the batch discount[1]

How developers use the Mistral OCR API

You call the OCR endpoint with a document URL, a base64 file or an uploaded PDF, and the model returns a JSON object[4]. Each page in it holds Markdown text, extracted images and tables, and optional header and footer fields. Tables come back as Markdown or HTML, set by a parameter[4]. For large jobs, Mistral points to its batch service, which halves the price[1].

Mistral also offers a Document AI playground in its AI Studio, where you drag in a PDF or image and see the parsed result[1]. That is the fastest way to see what the model reads before you write code.

Mistral OCR inside LoreAI

LoreAI calls Mistral OCR through the same API, using the mistral-ocr-latest alias, so every upload is read by Mistral's newest OCR model. The OCR result is not the end of the job. LoreAI translates the text into any of 105 languages, rebuilds headings, lists and tables in their original order, and returns an editable Word file.

1. Create a free account. Your first document is free, up to 3 pages and 3 MB.
2. Upload the scan or photo. A scanned PDF, a JPG or a PNG from your phone all work. LoreAI sends it to Mistral OCR to read.
3. Pick the target language. The source language is detected for you.
4. Download the Word file. You get a .docx with the translated text, the tables as Word tables, and names, dates and numbers as written.

Mistral OCR API and LoreAI, side by side

Which one should you use?

• Use the Mistral OCR API if you are building a product, a search index or a data pipeline and want Markdown or JSON at scale. The per-page price is low and the batch service halves it[1].
• Use the Document AI playground to test a few files by hand and see the raw output[1].
• Use LoreAI when you have a scan of a real document, such as a certificate, a transcript or a letter, and want it translated as an editable Word file without writing code.

The reading quality is the same model either way. LoreAI is the finished app around it, with translation, layout and the Word export done for you. For other OCR tools, see the OCR software comparison and the handwriting OCR comparison.

A machine translation is not a certified one

LoreAI is not a certified translation service. When an agency such as USCIS asks for a certified translation, a translator signs a statement that it is complete and accurate. The guide to getting a certified translation shows how to use LoreAI's Word file as the draft for that step.

Questions people ask