Mistral OCR 4

	Mistral OCR 4(mistral.ai)
	473 points by meetpateltech 23 hours ago \| 126 comments
	tl;dr: Mistral released OCR 4, a document extraction model that returns bounding boxes, typed block classifications (tables, equations, signatures, etc.), and per-word confidence scores across 170 languages, deployable in a single self-hosted container. It claims top scores on OlmOCRBench (85.20) and 72% win rates in human preference tests against competitors, though Mistral notes benchmark scoring artifacts inflate apparent errors on math and multi-column docs. Pricing is $4/1k pages via API ($2 batch), $5/1k for Document AI, available through Mistral Studio, AWS SageMaker, and Microsoft Foundry.
	HN Discussion: ↓Skepticism about Mistral's benchmark claims given past versions underperformed external benchmarks ↓Criticism of misleading presentation like truncated y-axes and selective benchmark reporting ↓Questioning the price increase and lack of clear differentiation from the previous v3 model ~Mixed real-world testing results showing accuracy issues with certain languages/scripts compared to competitors •Curiosity about missing competitor comparisons (e.g., Claude) in the benchmarks