Mistral OCR 4: Pricing, Bounding Boxes, and 170-Language Benchmarks Explained
- 3 hours ago
- 4 min read
Mistral OCR 4, released June 23, 2026 as the successor to OCR 3, returns extracted document text alongside paragraph-level bounding boxes, block-type classification, and confidence scores — structured output that plain OCR APIs don't include by default. It covers 170 languages across 10 language groups, processes up to 2,000 pages per minute on a single GPU, and topped the OlmOCRBench leaderboard with a score of 85.20 at launch.
At $4 per 1,000 pages, standard API pricing is exactly double what OCR 3 charged. For a team evaluating it, the deciding factor is whether bundled structured output — bounding boxes and block classification included in the base price — justifies paying more than double the $1.50-per-1,000-page rate that AWS, Google, and Azure charge for plain text extraction, given that those providers only add comparable structure at a separate, much higher pricing tier.
··········
WHAT OCR 4 RETURNS BEYOND PLAIN TEXT, AND WHERE IT RUNS.
Structured output, language coverage, and deployment channels define the product beyond the benchmark score.
OCR 4's output includes three elements a plain-text OCR call doesn't return: bounding boxes locating each text block on the page, classification of what kind of block it is (paragraph, table, figure caption, header), and a confidence score for each extraction. Confidence scores come in three granularities — page-level, word-level, or, as of the OCR 4.1 update on July 16, 2026, block-level — set via the confidence_scores_granularity parameter. That July update also means both the mistral-ocr-latest and mistral-ocr-4 model aliases now point to 4.1 automatically, so a pipeline calling either name picked up the block-granularity option without a code change.
Deployment isn't limited to Mistral's own API: OCR 4 is available through Mistral Studio, Amazon SageMaker, and Microsoft Foundry, and ships as a self-hosted single-container option for teams that need documents to stay inside their own infrastructure. It also integrates with the Mistral Search Toolkit, an open-source composable search framework, which feeds OCR 4's structured output directly into citation-ready chunks for RAG pipelines rather than requiring a separate parsing step before indexing.
........
Component | Detail | Function |
|---|---|---|
Structured output | Bounding boxes, block classification, confidence scores | Returns document structure, not just raw text |
Confidence granularity | Page, word, or block (block added in OCR 4.1, July 16, 2026) | Controls how finely confidence is reported per extraction |
Language coverage | 170 languages, 10 language groups | Single model handles multilingual document sets |
Throughput | Up to 2,000 pages/minute on one GPU | Sets expected processing speed for volume estimates |
Deployment | Mistral API, Studio, AWS SageMaker, Microsoft Foundry, self-hosted container | Matches deployment to existing cloud or residency requirements |
RAG integration | Mistral Search Toolkit | Feeds structured output into citation-ready chunks for retrieval |
........
··········
WHAT THE BENCHMARK CLAIMS ACTUALLY COVER, AND WHAT'S OUT OF SCOPE.
The 72% win rate and OlmOCRBench score come from a specific, disclosed evaluation setup — not a blanket accuracy claim.
Mistral's 72% win rate figure comes from blind human evaluation across more than 600 documents in 12-plus languages, judged head-to-head against competing OCR systems, with particularly strong results reported on multi-column layouts, tables with merged cells, and documents mixing text with figures — the layout types that trip up simpler extraction pipelines. Independent reporting checked the figure again roughly six weeks after launch and found it held up, which is more verification than most vendor-published benchmarks get, though it's still one evaluation design rather than a guarantee across every document type a given team will actually process.
Mistral's own documentation lists explicit out-of-scope uses: medical diagnosis, legal judgment, high-stakes financial decisions, safety-critical systems, real-time latency-sensitive paths, and non-document inputs like audio or video. That list matters more than it might read at first pass — it's an explicit statement that a 72% blind-eval win rate on document layout extraction says nothing about reliability in a workflow where the extracted text feeds a decision with legal or medical consequences, which is a different accuracy bar than getting the bounding boxes right.
··········
HOW $4 PER 1,000 PAGES COMPARES TO HYPERSCALER OCR PRICING.
Mistral's price sits between hyperscalers' plain-OCR tier and their structured-extraction tier — because it bundles what they charge extra for.
AWS, Google, and Azure all charge close to $1.50 per 1,000 pages for plain text extraction with no structure. None of them include bounding boxes or block classification at that price — getting comparable structured output means moving to a separate, materially more expensive tier on each platform.
........
Provider | Plain OCR | Structured extraction (tables/forms) |
|---|---|---|
Mistral OCR 4 | Included — $4/1,000 pages ($2/1,000 batch) | Included in the same price; Document AI mode $5/1,000 |
AWS Textract | ~$1.50/1,000 pages | Tables ~$15/1,000; Forms ~$50/1,000; Forms+Tables+Queries ~$70/1,000 |
Google Document AI | ~$1.50/1,000 pages | Layout Parser (tables) $10/1,000; Custom Extractor $30/1,000 |
Azure Document Intelligence | ~$1.50/1,000 pages | Custom extraction $30/1,000 |
........
The comparison that matters isn't OCR 4's $4 against the hyperscalers' $1.50 — it's OCR 4's $4 against the $10–$70 per 1,000 pages those same providers charge once structured output is actually needed. For any workload that requires bounding boxes, block types, or confidence scores rather than plain text, OCR 4 is priced well under every hyperscaler's structured tier; for a workload that only ever needed plain text, the hyperscalers' $1.50 rate remains cheaper, and paying for structure that goes unused is the mistake several reviewers flagged when teams switched from OCR 3 without checking which features their pipeline actually consumes.
··········
THE DECISION RULE FOR EVALUATING OCR 4.
The pricing comparison only holds if a workload actually consumes the structured output — bounding boxes, block classification, or per-block confidence scores feeding a downstream system, most commonly a RAG pipeline through the Search Toolkit integration. A pipeline that only needs plain extracted text, with no layout information and no confidence filtering, is better served by a hyperscaler's $1.50-per-1,000-page plain-OCR tier, since OCR 4's bundled structure goes unused at more than double the cost. A pipeline that would otherwise pay a hyperscaler's $10–$70 structured-extraction rate — invoices, multi-column reports, forms with merged cells — gets the same category of output from OCR 4 at a fraction of the price, plus 170-language coverage and self-hosting for teams that need documents to stay off third-party infrastructure entirely. The version-pinning detail is worth checking before deploying either way: calling mistral-ocr-latest or mistral-ocr-4 now silently resolves to 4.1, so a pipeline built and tested against one confidence-granularity behavior should confirm it still gets that behavior rather than assuming the alias stayed frozen.
·····
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····