Mistral Doubled Its OCR Price and Called It Progress

Buried in Mistral's OCR 4 announcement, between the benchmark charts and the partnership logos, is a number that nobody seems to have noticed: $4 per thousand pages, standard. OCR 3 was $2. They doubled the price. The announcement doesn't mention this.

That's actually the most interesting thing about OCR 4 — not the benchmarks, which don't hold up, and not the "SOTA" headline, which is wrong.

On OlmOCRBench, Mistral claims a "top score of 85.20." The public leaderboard disagrees. As of June 2026, Infinity-Parser2-Pro sits at 87.6% and Chandra-2 at 85.9% — both above Mistral's self-reported figure. Mistral OCR 4 doesn't appear on the leaderboard at all. The company appears to have evaluated against an earlier snapshot and then reported it as a current top score. Mistral's own announcement, to its credit, acknowledges that automated benchmarks have "significant limitations including ground-truth annotation errors" — which is a strange thing to say immediately after citing those same benchmarks as proof of superiority.

The 72% win rate in blind evaluation is Mistral's own human preference study on 600 documents. No independent replication exists. It may be accurate. It is not what "independently verified" looks like.

So: inflated benchmark claims, a self-undermining caveat, a price hike buried in the fine print. I'd normally stop there.

But here's the part I actually find interesting: the price hike is justified. Not the way Mistral tells it, but in the way that actually matters.

OCR 4 adds bounding boxes per word and per block, typed block classification (title, table, equation, signature, body text), and per-page confidence scores. These aren't accuracy improvements. They're a different product. OCR 3 gave you text. OCR 4 gives you a document graph — spatial coordinates, structural types, confidence weights — the primitives you actually need to build a RAG pipeline that doesn't chunk by character count and hope for the best. That's worth paying more for. AWS Textract charges $65 per thousand pages for comparable structured output. Mistral at $4 is still 16x cheaper.

The self-hosting option is the quieter competitive move. Google Document AI and AWS Textract are cloud-only. Mistral ships a single container you can run on-premise. For any regulated industry — banking, insurance, healthcare, government — cloud-only isn't an option for sensitive documents. Mistral knows this. That's the moat they're building.

Where does this go? I think structured document output — bounding boxes, block types, confidence scores — becomes the baseline expectation for production RAG pipelines within 18 months. "OCR" as a category gets replaced by "document intelligence," and the vendors who defined the output schema early have an advantage. Mistral is trying to be one of them.

The benchmark claims are embarrassing. The product is legitimate. I'd just prefer they were honest about which story they're actually in.