Mistral OCR 4

The useful read from the comments is that OCR is not one problem. Printed PDFs, bad phone photos, handwriting, historical scans, multilingual text, and layout reconstruction are all different workloads, and performance varies a lot by slice. People who had actually used Mistral’s earlier OCR models reported strong results on degraded archives, handwritten forms, and odd layouts, often saying modern vision-language models now beat older tools like ABBYY FineReader on messy inputs. At the same time, multiple people distrusted the launch numbers because Mistral had oversold past OCR releases with thin internal evals, and because the post dismisses public benchmarks like olmOCRBench and OmniDocBench as limited while showcasing internal results instead. Pricing got a more nuanced reaction than the headline suggests. Some called $4 per 1,000 pages cheap, others said Google Vision OCR is much cheaper, and the more grounded comparison was that Mistral is selling layout-aware document understanding rather than bare text extraction. That puts it closer to Azure Document Intelligence or Google’s richer document products than to classic OCR APIs. Several comments also drew a practical line between traditional OCR and model-based OCR. The newer systems can recover structure and handle ugly scans better, but they can also normalize punctuation, translate, or hallucinate text. That makes them powerful for ingestion pipelines and risky for exact transcription unless you keep a review loop. The strongest edge-case signal came from language coverage and image quality. One person testing Malayalam said normal handwriting worked but a slightly different style was misread as Kannada, while another pointed out Mistral initially labeled some language groups as “minor languages” before changing the wording. That sharpened the sense that multilingual claims should be treated carefully outside the languages vendors benchmark most. A few people were also surprised Claude was missing from the comparison set, though firsthand reports in the comments suggested Claude vision has lagged GPT and Gemini for OCR-oriented work. Overall mood was that Mistral may genuinely have a good OCR product, which is not true of every part of its lineup, but you should believe the customer eval before the marketing chart.