Mistral OCR 4 knows where each word lives and how much to trust it

Mistral AI today released OCR 4, the latest version of its document extraction model, adding bounding boxes, typed-block classification, and per-word confidence scores to its output. The model targets enterprise search, retrieval-augmented generation (RAG), and agentic workflows. It runs fully self-hosted on a single container. Microsoft's new platform gives scientists a governed…

Structured output beyond raw text

Older versions focused on converting pages into clean text and tables. OCR 4 returns a structured representation: every block gets a bounding box, a classification (title, table, equation, signature, and others), and inline confidence scores at the page and word level. Downstream systems know not just what a document says, but where each element sits, what role it plays, and how confident the model is in each region.

The structured output supports several workloads directly: semantic chunking for RAG, structural primitives for agents handling form filling or invoice processing, and consistent typed output for ingestion and indexing pipelines. IBM's new open-source agent framework cuts the…

Benchmark performance and human preference

Mistral OCR 4 scored 85.20 on OlmOCRBench and 93.07 on OmniDocBench, the top overall scores on both public benchmarks. The company also ran a head-to-head human evaluation on more than 600 documents across 12-plus languages, sourced from third-party vendors. Independent annotators preferred OCR 4 over every competing system, with win rates averaging 72%. Ifbench reveals the instruction-following gap that…

Mistral noted known limitations in automated benchmarks. During an audit of mismatches behind their scores, most turned out not to be model errors but artifacts of how benchmarks compare output: ground-truth errors, equivalent math notation counted as mismatches, equation segmentation issues, multi-column reading order assumptions, and block-type attribution problems. Mistral treats the aggregate benchmark scores as directional rather than definitive.

Multilingual coverage and performance

OCR 4 supports 170 languages across 10 language groups. On Mistral's internal Crawl Multilingual evaluation, the model leads across all eight language groups tested: English, Western Europe, Eastern Europe, Middle Eastern, Chinese, East Asian, Southeast Asian, and specialized languages. The gap is widest for specialized and low-resource languages, where many competing systems degrade sharply while OCR 4 maintains high accuracy.

Deployment options and pricing

The model runs on a single container, letting organizations with data-sovereignty requirements keep document data in their own infrastructure. Mistral OCR 4 through the API costs $4 per 1,000 pages, with a 50% Batch-API discount that drops it to $2 per 1,000 pages. Document AI, a no-code layer that feeds OCR output through a smaller model to generate structured JSON, costs $5 per 1,000 pages. Ai2 just opened an AI cluster that publicly shares…

OCR 4 is available via API through Mistral Studio, Amazon SageMaker, and Microsoft Foundry. Snowflake Parse Document support is coming soon. How alibaba cloud pushed its way into 20 gartner…

Integration with Search Toolkit

OCR 4 is an ingestion component of Mistral's Search Toolkit, an open-source, composable search framework announced at the AI Now Summit. Its structured output supplies citation-ready inputs to the toolkit's retrieval and evaluation workflow for RAG and enterprise search.

Early user feedback

Aidan Donohue, AI Engineer at Rogo, said OCR 4 achieved equivalent accuracy to leading agentic document parsers on a chart- and figure-dense financial QA dataset at roughly 8x lower cost and 17x lower latency. Ivan Mihailov, AI engineer at Anaqua, reported that Mistral OCR is roughly 4x faster per page than their incumbent provider, a result he called impressive for high-volume docketing workflows where speed is critical. The verification horizon: why verifying coding agents…