Top AI Product

Mistral dropped OCR 4 on June 23, an open-weights document-understanding model that does more than turn PDFs into text. It returns bounding boxes (where each block sits), block classification (title, table, equation, signature), and inline confidence scores per page and per word. So when the model isn’t sure about a line in a scanned invoice, it says so — instead of quietly hallucinating a number into your pipeline.

That confidence signal is the whole point. Document OCR is the front door for agents reading enterprise data, and silent errors there poison everything downstream.

Why it matters

The numbers back it up: independent annotators preferred OCR 4 over every leading OCR system tested, averaging a 72% win rate, plus the top score on OlmOCRBench (85.20). It handles 170 languages, eats PDF, DOC, PPT and OpenDocument, and is compact enough to self-host in a single container — keeping sensitive documents inside your own walls.

API and use cases

It ships through the Mistral API at $4 per 1,000 pages, halving to $2 via the batch discount. Self-hosting is available to enterprise customers. The typical job: feed messy invoices, contracts and forms into a RAG or agent pipeline and get back structured, citation-ready, position-aware output. Paired with the same-day Mistral Connectors launch, the agent last mile is getting filled in.

You Might Also Like

• Mistral Voxtral tts Scores 63 Listener Preference Over Elevenlabs and the Weights are Free
• Notion mcp Scores Product Hunt 1 With 408 Votes the Moment Mainstream Saas Joined the ai Agent Race
• Memvid Packs ai Agent Memory Into a Single File and Outperforms Sota rag by 35
• Microsoft Agent Governance Toolkit Scores 10 10 on Owasp Agentic Risks at 0 1ms per Check
• Mistral Medium 3 5 Scores 77 6 on swe Bench one Point shy of Gemini 3 1 pro