Mistral OCR 4.1 reaches GA: structure-aware parsing at $4/1K pages
The stable release returns bounding boxes, typed blocks and confidence scores in 170 languages — pricier than Cohere's Parse 5, but built for RAG citations.
Copy markdown
Preview to GA, safe to build on now
`mistral-ocr-4-1` graduated from public preview — live since July 16 — to general availability on Aug 31, and is now the recommended stable model ID. GA means a production SLA and no preview endpoint set to vanish under your pipeline.
$4 per 1,000 pages, or $2 in batch
The standard API is $4 per 1,000 pages, the non-real-time batch endpoint halves that to $2, and schema-based annotation runs $5. That undercuts no one — Cohere's Parse 5 shipped Aug 27 at $1.50/1K — so the premium buys structure, not a lower bill.
Bounding boxes make it citation-ready for RAG
Every block returns with a paragraph-level bounding box, a typed label (title, table, equation, signature) and a block-level confidence score, across 170 languages. Where Parse 5 hands you markdown, OCR 4.1 tells you where on the page each fact sits — the missing piece for grounded citations in RAG and agentic document workflows.
Self-host the whole model in one container
Enterprise customers can run a single-container self-hosted deployment for documents that can't leave the building (via sales). Human raters preferred OCR 4 on 72% of documents against rival systems, though handwriting still trails Claude.
Elsewhere: Google retires Gemini Robotics-ER 1.6 Preview
The `gemini-robotics-er-1.6-preview` endpoint shut down at the end of Aug 31 — migrate to `gemini-robotics-er-2-preview`. If you're prototyping embodied agents on Gemini, check your hardcoded model IDs before calls start 404-ing.