Mistral launches an OCR API that turns any PDF into Markdown at $1 per 1,000 pages
Mistral OCR is a multimodal document-understanding endpoint that takes PDFs and images and returns ordered, interleaved text and image output formatted as Markdown — preserving tables, LaTeX equations, and layout rather than flattening a page into a wall of characters. Mistral reports roughly 94.9% overall accuracy on internal benchmarks against GPT-4o at 89.77%, Gemini 2.0 Flash at 88.69%, and Azure OCR at 89.52%, processes up to 2,000 pages per minute per node, and prices at $1 per 1,000 pages with roughly double the pages per dollar under batch inference.
The distinction that matters is architectural: this is a vision-language model trained for document understanding, not character-by-character OCR, which is why it can place bounding boxes around illustrations and emit Markdown that a RAG system can ingest directly. Mistral made it the default document model for Le Chat and says it outperforms Google, Microsoft and OpenAI APIs on complex documents and non-English text. It is not open source — there are no public weights, and self-hosting is offered selectively for sensitive deployments.
The launch-day reaction was the useful part. On Hacker News the release drew 1,756 points, and alongside the enthusiasm came concrete failure reports: users documented hallucinated non-existent words and repeated text blocks in French and English documents, with the errors described as hard to spot. In a domain where faithfulness is the entire product, an LLM-based parser that confidently invents text is a different kind of risk than a chatbot that does.
WHY IT MATTERS
This is the retrieval layer getting a real price. If you run RAG over contracts, filings, slide decks or manuals, $1 per 1,000 pages with Markdown output removes a hand-rolled parsing stack from the critical path — and it sets a public benchmark for what the closed labs must beat. The hallucination caveat is the part to design around: for anything where an invented sentence is a liability, the parser output needs its own verification step, because the failure mode is plausible text rather than an error code.
