June 22, 2026: Baidu open-sources Unlimited-OCR under MIT — one-shot multi-page document parsing, free forever, covered in today’s dispatch.
June 23, 2026: Mistral ships OCR 4.
One day apart. And here is the honest correction to the framing this column almost ran with: that is not a counterpunch. Model launches are locked months in advance — nobody rebuilds a release overnight. What actually happened is more telling than a reaction would be: the document-AI release cadence is now so dense that a European commercial lab’s planned flagship lands within 24 hours of a Chinese open-weight release in the same category, with neither party reacting to the other. The punches aren’t being traded. They’re just raining.
What the two releases say side by side is the real story.
24 hours apart. Nobody reacted.
That’s the point.
Baidu open-sources Unlimited-OCR on June 22. Mistral ships OCR 4 on June 23. Not a counterpunch — launches are planned months out. The cadence is now so dense that two roadmaps collide within a day — and their pricing tells opposite stories.
One category, one day, two theories
Nearly tied on the shared yardstick, priced a universe apart — because they’re not selling the same thing.
The ladder that runs the wrong way — on purpose
Per 1,000 pages, list price. While the open floor fell to zero, Mistral doubled its price twice — repricing upward into the layer free models don’t ship. That’s a company that read the memo precisely.
What each side actually sells
The $0 tier ships
- Transcription: pages → markdown, weights yours
- Sovereignty: run it, own it, keep it
- Zero marginal cost at any volume
The $4 tier ships
- Structure: bounding boxes, typed blocks, per-element confidence, schemas
- Jurisdiction: self-hosted single container — in your building, but not open weights; the license bill still arrives
- Accountability: SLA, contract, someone to blame
The 93.07 OmniDocBench and 72% win-rate figures are vendor-stated; on the public OlmOCRBench leaderboard (May 21 update), OCR 4 would place roughly third — not first. Third on a contested public board is a strong model. Launch pages are launch pages — a rule applied to Baidu’s numbers too.
Also reported, not confirmed: Mistral targeting €1B 2026 revenue (from ~€200M), early talks near €3B at ~€20B valuation. Document AI is a layer that revenue has to come from.
ScanSnap iX2500 Wireless or USB High-Speed Cloud Enabled Document, Photo & Receipt Scanner with Large 5" Touchscreen and 100 Page Auto Document Feeder for Mac or PC, Black
- Advanced Touchscreen: Large 5-inch touchscreen for easy control
- High-Speed Scanning: 45 pages per minute double-sided scanning
- Large Document Capacity: 100-page automatic document feeder
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Two products, one benchmark, opposite theories
Unlimited-OCR’s theory: transcription is the product. Pages in, markdown out, weights free, run it yourself.
Mistral OCR 4’s theory: transcription is the commodity — structure is the product. Per Mistral’s launch material: paragraph-level bounding boxes, typed block classification, per-element confidence scores, 170 languages, a schema-driven Document AI mode, and a self-hosted single-container option for keeping documents inside your jurisdiction. Priced at $4 per 1,000 pages, $2 in batch, $5 for Document AI — through an API, with an SLA, from a company you can put in a contract.
On the shared yardstick they’re nearly tied: Mistral states 93.07 on OmniDocBench against Unlimited-OCR’s 93.23. Read the Mistral number with the standard discount — it’s vendor-stated, and analysis of the launch noted that on the public OlmOCRBench leaderboard (as of its May 21 update), OCR 4 would place roughly third, not first, despite topping Mistral’s own chart. The blind-human-evaluation “72% win rate” is likewise Mistral’s own study. None of this means the model is weak — third on a contested public leaderboard is a strong model. It means launch pages are launch pages, a rule this site applies to Baidu’s numbers too.
Ambir TravelScan Pro 600-BCS Simplex Document Scanner with AmbirScan Business Card for Windows PC and MAC
- Easy Outlook Contact Import: Scan business cards directly into Outlook
- Duplicate Contact Management: Automatically detect and delete duplicates
- AI-Powered Data Recognition: Enhanced accuracy with AI processing
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The pricing ladder that runs the wrong way — on purpose
Here’s the detail worth a moment of genuine appreciation. Mistral’s OCR pricing history: the original Mistral OCR (March 2025) at $1 per 1,000 pages. OCR 3 at $2 ($1 batch). OCR 4 at $4 ($2 batch).
While the open-weight floor fell to zero, Mistral doubled its price twice.
That is not a company that missed the memo — it’s a company that read it precisely. When free 3B models commoditize transcription, you don’t compete at the commodity layer; you reprice upward into the structure layer the free models don’t ship: bounding boxes, typed extraction, confidence, jurisdiction-contained deployment, and someone to blame. A vendor-published customer quote from the launch (Rogo, a financial-AI firm) claims equivalent accuracy to agentic parsers at roughly 8× lower cost and 17× lower latency — vendor-published, but the shape of the claim is the strategy: sell the workflow, not the words.
The context makes the bet legible: reporting around the launch places Mistral targeting €1B revenue in 2026 (from ~€200M) and in early talks for a raise near €3B at roughly a €20B valuation — reported figures, not confirmed. Document AI is one of the layers that revenue has to come from, and it’s a layer where “European, self-hostable, contractable” is a real moat against both American clouds and Chinese open weights.
As an affiliate, we earn on qualifying purchases.
The honest ledger
What holds up: the self-hosted container is a genuinely differentiated answer for EU-regulated buyers who need cloud-grade structure without third-country processing — with the standing correction that self-hosting a commercial model is not open weights: you run it in your building, you don’t own it, and the license bill arrives either way. For buyers whose requirement is jurisdiction, that’s fine. For buyers whose requirement is sovereignty, it isn’t, and the two get conflated in every sales deck.
What to watch: the squeeze works both directions. Tuesday’s dispatch showed 0.9B open specialists at 94.5+ on the same benchmark — the accuracy gap justifying a paid API narrows by the month, which is precisely why Mistral’s value proposition keeps migrating up-stack, away from where the free models are. Whether structure-layer features stay ahead of what the open ecosystem wraps around free transcription models is the actual race. This pipeline’s Stage 4 — schema extraction with a local LLM — is exactly that wrapper, built in an afternoon.
Monday’s thesis piece will name this pattern across the whole week: commodity below, structure above, and the frontier labs repositioning in real time. Mistral, to its credit, repositioned before the floor arrived — with a launch date that proves the releases now come too fast for anyone to be reacting to anyone.
As an affiliate, we earn on qualifying purchases.
Sources
- Mistral AI, “Mistral OCR 4” launch announcement (June 23, 2026) — feature set, $4/$2/$5 pricing, self-hosted container, 170 languages, OmniDocBench 93.07 and win-rate figures (vendor-stated), Rogo customer quote (vendor-published)
- Mistral AI, “Mistral OCR” (March 2025) and “Introducing Mistral OCR 3” — $1 and $2/$1 pricing history
- Digital Applied, “Mistral OCR 4: Document AI for Business Automation” (June 2026) — public OlmOCRBench placement analysis (~third as of May 21 update); €1B 2026 revenue target and reported ~€3B raise at ~€20B (reported, unconfirmed); Azure price comparisons
- AI Made Tools, “Mistral OCR 4: Complete Guide” (June 2026) — competitive positioning vs Google Document AI and Baidu Unlimited-OCR
- Baidu, Unlimited-OCR model card and arXiv:2606.23050 (June 22–23, 2026) — release date and 93.23 comparison basis
- ThorstenMeyerAI.com, “Baidu’s Unlimited-OCR Reads a 40-Page PDF in One Pass” (July 21, 2026) — companion dispatch
Vendor-stated benchmarks, vendor-published testimonials, and reported financials are labeled as such throughout. The one-day launch proximity is treated as roadmap coincidence, not reaction — and that is the point.
