What's this about?
Mistral, the French AI lab, released a new version of its OCR model, which converts images of documents into machine-readable text. The HN reception was notably cold, centering almost entirely on price: at roughly €3.5 per 1,000 pages, commenters compared it unfavorably to self-hosted alternatives and to Chinese models.
What exactly is OCR, and why isn't it solved?
OCR (optical character recognition) turns a picture of text into actual text you can search and edit. It's been around for decades and works well on clean, modern, printed pages. It stays hard on everything else: historical typefaces like Fraktur, ligatures, mathematical subscripts and superscripts, multi-column layouts, handwriting, and scholarly notation. Modern approaches use vision-capable language models, which handle messy layouts far better than older systems but cost dramatically more per page — which is exactly the trade-off this thread argued about.
What's the community saying?
- The price comparison dominated. One commenter running an OCR pipeline on rented GPUs reported roughly $0.05–0.10 per 1,000 pages at 0.8 seconds per page with full bounding boxes, calling $3.5 per 1,000 "just too expensive."
- Speed is the genuine defense. Someone with internal benchmarks confirmed Mistral OCR is significantly faster than comparable APIs, even without commenting on accuracy.
- It fails on hard documents. A user who tests every new release against a scanned book with ligatures, critical sigla and Fraktur letterforms reported nothing special here, and said OpenAI's pro models still dominate for detailed work.
- The Europe question. Two of the most-upvoted comments weren't technical at all — "the Chinese did it better, Mistral is alive thanks to regulations" and a broader lament about Europe's role in the AI race.
- Local models are winning this niche. Several reported good results from NuExtract or Baidu's OCR running on a single consumer GPU, beating Tesseract and generic vision models at modest throughput.