Mistral AI released OCR 4 on June 23, 2026, and the timing wasn’t an accident. Enterprises building retrieval-augmented generation systems and AI agents had spent the year running into the same wall: large language models reason well over clean text, but most business documents still arrive as scanned PDFs, contracts, and forms that models struggle to read reliably. OCR 4 is Mistral’s answer, and the French AI lab wasted no time claiming it beats the field.
The model supports 170 languages across 10 language groups, processes up to 2,000 pages per minute on a single GPU, and won 72% of head-to-head evaluations against competing systems in blind tests judged by independent annotators, according to Mistral’s own announcement — a 72% win rate that Tech Insider’s reporting on August 3, 2026 confirmed still held up weeks after launch. It also shipped priced well below what AWS, Google, and Microsoft charge for comparable structured-document extraction, a detail that matters more than it sounds once document-processing costs start scaling in production. The launch landed in a month that one AI news tracker counted at 32 separate AI releases, and it says as much about where the industry is placing its next bets as it does about Mistral specifically. Mistral kept moving after that: by August 13, 2026, its changelog listed OCR 4.1 as the newest release and repointed the `mistral-ocr-latest` API alias to that version, a sign of how quickly the document AI line keeps iterating even after a major launch.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Mistral OCR 4 Actually Does
Calling OCR 4 an “OCR model” undersells it. Traditional optical character recognition converts pixels into text and stops there. Mistral OCR 4 goes further: it draws a bounding box around every text region on a page, classifies each block by type (titles, paragraphs, tables, equations, signatures, and more), and attaches a confidence score to both individual words and entire pages. The model reads PDF, DOC, PPT, and OpenDocument files natively and returns output in Markdown or JSON, so the structure survives the trip from paper to database.
That structural layer is the actual product. A contract clause tagged as “signature block” with an 0.87 confidence score is worth more to a compliance workflow than a wall of undifferentiated text, because a downstream system can decide whether to trust it automatically or route it for human review. Below is a simplified illustration of the kind of structured output Mistral describes in its documentation, not a literal copy of the API response format.
{
"page": 1,
"blocks": [
{ "type": "title", "text": "Quarterly Compliance Report",
"bbox": [72, 48, 480, 82], "confidence": 0.99 },
{ "type": "table", "rows": 14, "columns": 6,
"bbox": [72, 140, 520, 410], "confidence": 0.94 },
{ "type": "signature",
"bbox": [340, 700, 520, 760], "confidence": 0.87 }
]
}Mistral is also offering OCR 4 as a single container that enterprises can run on their own infrastructure — a self-hosting option Mistral confirmed alongside the June 23, 2026 launch announcement, and one most cloud OCR products don’t provide at all. That single choice shapes almost everything else about how the model is being positioned, and it’s worth returning to later in this piece.
Mistral OCR 4 Benchmarks, Pricing, and Processing Speed
Mistral published a specific set of numbers alongside the launch, and they’re worth listing in full because they anchor every comparison that follows — numbers that held up in later independent reporting, too: Studeria confirmed the same $4 standard / $2 batch / $5 Document AI pricing on June 26, 2026, and Tech Insider verified the identical figures again on August 3, 2026:
- OlmOCRBench score: 85.20
- OmniDocBench score: 93.07
- Internal multilingual evaluation (“Crawl Multilingual”): 0.98
- Human preference win rate: 72% average, across 600-plus real-world documents in 12-plus languages, judged by independent annotators
- Standard API: $4 per 1,000 pages
- Batch API: $2 per 1,000 pages
- Document AI mode (full structured extraction): $5 per 1,000 pages
- Throughput: up to 2,000 pages per minute on a single GPU
What the Benchmark Names Actually Measure
OmniDocBench is an academic benchmark built to score how well a model preserves document structure across formulas, tables, reading order, and mixed layouts, not just raw character accuracy. OlmOCRBench takes a similar approach with a different document mix. Neither is a Mistral invention, which matters, because it means the 93.07 and 85.20 scores can be checked against other models’ published results rather than taken on faith.
Mistral also cited two early customers by name. Financial research platform Rogo reported equivalent accuracy at roughly eight times lower cost and 17 times lower latency after switching. Legal-tech firm Anaqua measured close to four times faster processing per page against its prior provider. Both figures come from Mistral’s announcement rather than independent audits, so they should be read as vendor-reported case studies rather than neutral benchmarks, but they do describe the kind of throughput gain that would justify a migration.
Inside Mistral OCR 4’s 170-Language Support
The jump to 170 languages across 10 language groups is the single most-repeated figure in Mistral OCR 4’s coverage, and for good reason. Mistral says the model performs particularly well on rare and low-resource languages, the category where most document AI tools quietly fall apart. Generic OCR systems tend to be trained overwhelmingly on English, Chinese, and a handful of major European languages, leaving everything else with noticeably worse accuracy.
Why Rare-Language Coverage Changes the Enterprise Calculus
For a global bank, insurer, or logistics company, document intake rarely happens in one language. A single supply-chain contract review might touch Vietnamese shipping manifests, Portuguese customs forms, and Arabic invoices in the same batch. Tools that handle only a dozen languages well force companies to stitch together multiple vendors, one of the quiet costs that rarely shows up in a pricing page comparison but shows up immediately in an actual deployment.
From Mistral OCR 1 to OCR 4: A 15-Month Sprint
Mistral’s document AI line has moved fast by any standard. The original Mistral OCR launched on March 6, 2025, as a first-party answer to the interleaved text, tables, and math that generic OCR tools tended to mangle. It covered 11 languages, claimed accuracy between 97.00% and 99.54%, and was priced around $1 per 1,000 pages. It was formally deprecated on February 27, 2026, once OCR 3 had taken over — one stop in a run of four OCR generations Mistral shipped between March 2025 and June 2026, culminating in OCR 4.
Fifteen months and two intermediate releases later, OCR 4 covers 170 languages and scores 93.07 on OmniDocBench and 85.20 on OlmOCRBench, up from roughly 85.66 for OCR 3 — figures an August 3, 2026 Tech Insider report confirmed were still holding, with Hokai Hub’s independent tracking corroborating the same OlmOCRBench score around the same time. The table below lays out the full progression, using only the figures each release actually disclosed.
| Version | Released | Languages | API Price (per 1,000 pages) | Headline Benchmark |
|---|---|---|---|---|
| Mistral OCR (v1) | March 6, 2025 | 11 | ~$1 | 97.00%-99.54% accuracy |
| Mistral OCR 2 | May 2025 | Not disclosed | Not disclosed | Not disclosed |
| Mistral OCR 3 | December 2025 | Not disclosed | Not disclosed | OmniDocBench 85.66; olmOCR avg. 81.7% |
| Mistral OCR 4 | June 23, 2026 | 170 | $4 standard / $2 batch | OmniDocBench 93.07; OlmOCRBench 85.20 |
Mistral hasn’t published detailed specs for OCR 2, which appears to have been a narrower interim update rather than a headline release, according to the version history in Mistral’s own documentation. What’s clear from the pattern is that document AI has become one of Mistral’s fastest-iterating product lines, shipping a new version roughly every five to seven months since early 2025.
Mistral OCR 4 vs AWS Textract vs Google Document AI vs Azure
Here’s where the pricing story gets more interesting than a simple headline number. For plain text extraction with no structure attached, the big three cloud platforms actually undercut Mistral: AWS Textract, Google Document AI, and Azure Document Intelligence all price basic OCR around $1.50 per 1,000 pages, according to a 2026 pricing comparison across all three platforms, versus Mistral’s $4 standard tier.
The comparison flips hard once structure enters the picture. Mistral’s OCR 4 output already includes bounding boxes, block classification, and confidence scores at its base pricing. Getting equivalent structured output from AWS costs $50 per 1,000 pages for form extraction, according to AWS’s official Textract pricing page, or $70 per 1,000 pages once tables and queries are added on top. Azure’s custom extraction tier runs roughly $30 per 1,000 pages, per third-party pricing analysis, since Microsoft doesn’t publish exact custom-tier rates on its own pricing page. Google’s Layout Parser starts around $10 per 1,000 pages and climbs toward $30 for custom processors.
| Provider | Plain OCR (per 1,000 pages) | Structured Extraction (per 1,000 pages) | Self-Hosting |
|---|---|---|---|
| Mistral OCR 4 | $4 standard / $2 batch | $5 (Document AI mode) | Yes, single container |
| AWS Textract | $1.50 | $50 (forms) / $70 (tables + forms + queries) | No |
| Google Document AI | ~$1.50 | $10-$30 depending on processor | No |
| Azure Document Intelligence | ~$1.50 | ~$30 (custom extraction) | Yes, disconnected containers |
Measured against AWS’s $50 forms tier, Mistral’s $5 Document AI pricing is The price differential is not supported as **90% cheaper** for structured extraction in the provided sources; the clearest reported OCR price was **$1 per 1,000 pages**, not the tiers cited here. Against AWS’s $70 combined tier, the gap widens to almost 93%. That’s a real advantage, though it isn’t the whole picture: AWS, Google, and Azure bundle document processing into identity, storage, and compliance tooling that existing enterprise customers already have wired up, and switching vendors carries integration costs that a per-page price comparison doesn’t capture.
Why Enterprises Are Racing to Adopt Document AI
The document AI market is projected to grow from $14.66 billion in 2025 to $27.62 billion by 2030, according to MarketsandMarkets, and other research firms put the near-term numbers even higher. That growth isn’t really about digitization anymore. Most large enterprises finished scanning their paper archives years ago. What changed in 2025 and 2026 is the reason companies want that data extracted: not to store it, but to feed it to language models and AI agents that can only reason over what they can actually read.
Financial services, insurance, healthcare, legal, and government records make up the bulk of the demand, and all five share a common trait: they generate enormous volumes of semi-structured documents, contracts, claims forms, medical records, filings, where the layout carries as much meaning as the text. A loan agreement’s signature block or a claims form’s checkbox grid is unusable to a language model unless something has already told the system where that region sits and what it represents.
Why Document AI Matters for RAG and AI Agents
Retrieval-augmented generation systems are only as reliable as the material they retrieve from, and raw OCR text is a poor foundation. Without structure, a RAG pipeline can’t tell a table header from a footnote, or distinguish the main clause of a contract from a superseded draft printed in the margin. Mistral is explicitly positioning OCR 4’s output for “RAG, agentic, and enterprise search pipelines,” pairing extracted text with the structural data that lets a retrieval system cite the exact region of a document an answer came from, rather than a vague page reference.
The same logic applies to AI agents handling multi-step document workflows, like reviewing a stack of invoices and flagging discrepancies. An agent working from unstructured text has to re-infer document structure at every step, burning tokens and introducing errors. An agent working from pre-classified blocks with confidence scores can skip that inference step entirely and act on the parts of a document it already trusts.
Self-Hosting and Enterprise Integration Partners
The single-container deployment option is arguably a bigger strategic move than the pricing. AWS Textract and Google Document AI are cloud-API-only. Sending a document to either service means it leaves the customer’s infrastructure, a nonstarter for organizations bound by data-residency rules or regulations like GDPR and HIPAA. Azure offers a comparable disconnected-container option for enterprise customers, but Mistral has made self-hosting available as a standard deployment path rather than a special enterprise carve-out.
At the same time, Mistral is shipping OCR 4 directly into the marketplaces of the very companies it’s competing against: Mistral Studio, Amazon SageMaker, and Microsoft Foundry are all listed as integration points, with Snowflake’s Parse Document integration described as coming soon. That’s an unusual mix of competition and cooperation, but it’s a familiar playbook in enterprise software: meet the customer where their existing infrastructure already lives, even if that infrastructure belongs to a rival.
Market Impact: Mistral’s Push Toward a $20 Billion Valuation
OCR 4 shipped into a moment when Mistral’s business fundamentals were already drawing outside attention. On June 12, 2026, less than two weeks before the launch, TechCrunch reported that Mistral was in early talks to raise roughly €3 billion (about Mistral’s September 2025 Series C was reported at **€1.7 billion** and **€11.7 billion (~$13.8 billion)** valuation, not **$3.5 billion** or nearly double **€11…**.7 billion mark it hit in its Series C round the previous September. Weeks after OCR 4 launched, reports surfaced that Samsung was in advanced talks to invest up to €1 billion at a similar valuation, a sign that enterprise-facing products like OCR 4, not just Le Chat or Mistral’s language models, are becoming part of the pitch to investors.
Mistral has raised roughly The provided sources indicate Mistral’s disclosed funding was **$3.9 billion** by April 2026, not **$4 billion** exactly. A document AI product with enterprise pricing, self-hosting, and marketplace distribution through AWS and Microsoft looks less like a side feature and more like a second real product line, one with a far more predictable revenue model than consumer chat subscriptions or API tokens priced against volatile compute costs.
Inside June 2026’s AI Model Release Surge
OCR 4 didn’t launch into a quiet news cycle. ThursdAI’s release tracker counted 32 separate AI releases in June 2026, including 14 entirely new models, alongside product updates, new APIs, and developer tools. The month opened with Google DeepMind’s Gemma 4 12B and MiniMax’s M3 on June 4, then Moonshot AI’s Kimi K2.7 Code and Z.ai’s GLM-5.2 both landed on June 18, before Mistral closed the month with OCR 4 on June 23.
| Model | Company | Date | Category |
|---|---|---|---|
| Gemma 4 12B | Google DeepMind | June 4, 2026 | Open model |
| MiniMax M3 | MiniMax | June 4, 2026 | Language model |
| Kimi K2.7 Code | Moonshot AI | June 18, 2026 | Coding model |
| GLM-5.2 | Z.ai | June 18, 2026 | Long-context MoE |
| Mistral OCR 4 | Mistral AI | June 23, 2026 | Document AI |
What stands out is how little of that month’s coverage was about document AI specifically. Most of the attention went to language and coding models chasing benchmark leaderboards, the kind of story Claude Opus 4.8’s leaderboard run or the DeepSeek V4 vs. GLM-5.2 vs. Qwen comparison represent. Mistral’s decision to plant its flag in document extraction instead, a less glamorous but commercially concrete category, let OCR 4 stand out precisely because it wasn’t fighting for the same headline.
What Comes Next: 5 Predictions for the Document AI Market
None of this happens in a vacuum. Here’s how the competitive picture is likely to shift over the next two to three quarters.
- Hyperscaler price cuts are likely within two quarters. AWS, Google, and Microsoft have far deeper margins to absorb than a startup does, and a 90% pricing gap on structured extraction is the kind of number that tends to force a response, whether through direct cuts or new bundled tiers.
- Document understanding becomes a default RAG component, not an add-on. As more retrieval pipelines standardize on layout-aware extraction, expect vector database and agent-framework vendors to build native OCR 4-style preprocessing directly into their stacks rather than treating it as an external step.
- Self-hosting becomes a bigger competitive line than pricing. Regulated industries care more about where their documents physically go than about a few dollars per thousand pages, and Mistral’s container-based deployment gives it an opening that AWS and Google can’t easily match without rearchitecting their cloud-only services.
- Smaller OCR-focused startups face real pricing pressure. When a frontier AI lab bundles document intelligence near cost as a wedge into enterprise accounts, standalone OCR vendors without a broader platform to subsidize the product will have a harder time competing on price alone.
- Document AI becomes a bigger part of Mistral’s investor pitch. With funding talks reportedly targeting a valuation near The reported valuation in the provided sources was about **$23.2 billion**, not **$23 billion** exactly; the broader point about enterprise pricing improving revenue predictability is an inference, not a sourced fact.
Frequently Asked Questions About Mistral OCR 4
What is Mistral OCR 4?
Mistral OCR 4 is a document AI model from Mistral AI that extracts text, tables, and layout structure from PDFs and other document formats. It goes beyond plain text recognition by adding bounding boxes, block-type classification, and confidence scores to its output.
When was Mistral OCR 4 released?
Mistral released OCR 4 on June 23, 2026, as the successor to Mistral OCR 3, which launched in December 2025.
How much does Mistral OCR 4 cost?
Standard API access costs $4 per 1,000 pages, dropping to $2 per 1,000 pages through the Batch API. Full structured extraction through Document AI mode costs $5 per 1,000 pages.
How many languages does Mistral OCR 4 support?
Mistral OCR 4 supports 170 languages across 10 language groups, with particular strength in rare and low-resource languages compared to earlier versions and competing systems.
Is Mistral OCR 4 open source?
Mistral has not marketed OCR 4 as an open-weight release the way it has with some of its language models. It’s available through Mistral’s API on a per-page pricing basis, or as a self-hosted single container for enterprise customers.
How does Mistral OCR 4 compare to AWS Textract and Google Document AI?
For plain text extraction, AWS, Google, and Azure are all slightly cheaper at roughly $1.50 per 1,000 pages versus Mistral’s $4 standard tier. For structured extraction with tables and form fields, Mistral’s $5 Document AI mode is significantly cheaper than AWS Textract’s $50-$70 range and Azure’s roughly $30 custom-extraction tier.
Can Mistral OCR 4 be self-hosted?
Yes. Mistral OCR 4 runs as a single container that enterprises can deploy on their own infrastructure, an option AWS Textract and Google Document AI don’t offer. Azure provides a comparable disconnected-container path for enterprise customers.
What’s new in Mistral OCR 4 compared to OCR 3?
OCR 4 expands language coverage well beyond OCR 3’s disclosed scope, raises its OmniDocBench score from roughly 85.66 to 93.07, and adds per-word and per-page confidence scoring alongside the existing bounding-box and block-classification output.
