According to Accord Beating monitoring, Mistral AI has released the Mistral OCR 4 document intelligent understanding model. In addition to text extraction, the model also outputs bounding boxes, element classification (such as titles, tables, formulas, signatures), and word-level and page-level confidence scores, supporting mainstream formats such as PDF, DOC, and PPT.
OCR 4 scored 85.20 on the multi-column and math formula benchmark OlmOCRBench, ranking first in the industry, and achieved a score of 93.07 on OmniDocBench. Early enterprise user tests at Rogo showed that in dense chart financial Q&A, the new model's parsing accuracy rivals that of the cutting-edge Agentic parser, at only one-eighth of the cost, with latency significantly reduced to one-seventeenth. The model now supports single-container self-hosting to ensure enterprise data privacy locally.
Standard extraction is priced at $4 per thousand pages, which can be reduced to $2 using the Batch interface. By opting for the $5 per thousand pages Document AI mode, users can input a custom JSON Schema to have back-end models like mistral-small-2603 output formatted data directly. The model is currently live on Mistral Studio, Amazon SageMaker, and Microsoft Foundry.
