Mistral OCR is a document OCR model from Mistral AI, a Paris-based AI lab, available through the /v1/ocr endpoint. The current version, OCR 4.1, was released on 16 July 2026 and is generally available at $4 per 1,000 pages, or $5 per 1,000 annotated pages when structured extraction is switched on (Mistral model card, accessed 13 October 2026).
It is a strong model at a very low price. If you are looking for a Mistral OCR alternative, anyformat solves an adjacent problem: schema-validated extraction wrapped in workflow orchestration, human review and European data sovereignty.
anyformat is enterprise-ready agentic document processing: agentic OCR that turns any document into data you can trust, where every value is validated and cites its source. It is built for companies with serious demands on security, scale and audit, so process owners and their IT teams use it alongside the developers who evaluate and integrate it.
You describe the fields once as a schema. Parse reads the document, Extract returns each value with a confidence score and its position on the page, Validate runs your rules, and low-confidence values go to a reviewer in the same workflow, through the API, the Python and TypeScript SDKs or the no-code Studio. Every value carries its evidence, it holds up on hard documents, it is a European company with ISO 27001 and GDPR, and the price covers the whole workflow, not only the extraction. Compared with a model call, the review, validation and orchestration come built in.
The measure that matters is how many documents go through without anyone touching them, and teams go live in days, not quarters, with SDKs, templates and evals to measure accuracy as they go.
What Mistral OCR is, and what it is not
Mistral OCR returns markdown plus paragraph-level bounding boxes, structural block labels and block-level confidence scores, all from the same endpoint. Batch processing runs on a separate /v1/batch endpoint. Mistral's OCR 4 launch post claims 170 languages across 10 language groups (Mistral, 23 June 2026).
It is one model in a broad catalog, not a document-processing product. The endpoint takes documents in and returns text and layout out; classification, splitting, validation, routing and review are not part of it. Teams that choose Mistral OCR build those steps around it.
Customization and extraction approach
Mistral's structured extraction is called annotation. The OCR markdown, plus the first eight images found on the document, is sent to a vision language model (mistral-small-2603) together with a JSON schema you define, and the model returns structured JSON (Mistral docs, accessed 13 October 2026). Schemas can be written in Pydantic, Zod or raw JSON Schema, and field descriptions act as instructions.
That is a single pass over the whole document. Mistral's documentation does not describe row-level extraction, cross-field checks or lookups against your own data.
anyformat extracts against a user-defined schema in zero-shot and returns validated JSON, with document-level and table-level fields declared together and returned in one call. Every field carries a confidence score and a visual citation pointing at its exact position on the page.
Confidence: reading the page versus trusting the value
OCR 4.1 adds block-level confidence to the word-level and page-level scores that OCR 4 introduced. These are off by default and must be requested per call. They tell you how well the model read a region of the page.
They do not tell you whether an extracted value is right. Mistral's annotation output carries no per-field confidence in its documentation, so a review process has nothing to route on at the field level.
anyformat returns a confidence per extracted field, with the source evidence alongside it, so a threshold can decide which values flow through and which go to a reviewer inside the same workflow.
For how the scores are produced and read, see how confidence scoring works.
Benchmark: where each one leads
ParseBench is an independent public benchmark of about 2,000 human-verified pages, published by LlamaIndex, which also sells a parsing product that tops its own leaderboard. On the leaderboard as read on 13 October 2026, anyformat (Standard) scored 80.30 overall, fifth of 147 entries. Mistral OCR 4 with annotations scored 68.23 and without them 60.68 (ParseBench leaderboard, 13 October 2026).
The category scores show the split:
Mistral reads text more faithfully and grounds it to the page better; anyformat is ahead on tables, charts and formatting. Two caveats apply. The leaderboard scored OCR 4, which Mistral deprecated on 29 September 2026 in favour of OCR 4.1, so the Mistral row is not the current version. And leaderboard scores measure parsing, not extraction against your own schema, which is why a test on your own documents is the real comparison.
anyformat also publishes its own parsing benchmark on 1,000+ real documents, with a different document set and scoring.
Accuracy: reading, structure and values
Accuracy means three different things in document AI, and the two products are strongest at different ones.
Reading the characters. This is transcription, and Mistral's evidence is strongest here. Its OCR 4 launch post reports a 72% average win rate in blind human preference tests on 600+ documents in 12+ languages, 85.20 on OlmOCRBench and 93.07 on OmniDocBench (Mistral, 23 June 2026). Mistral also says these numbers are directional, that both benchmarks have known scoring artifacts, that competitor scores are its own reproductions, and it recommends evaluating on your own documents. ParseBench agrees on this layer: Mistral scores 89.55 on content faithfulness against 87.35 for anyformat.
Reconstructing structure. This covers tables, charts and reading order. ParseBench puts anyformat ahead on tables (86.46 against 73.94) and charts (85.56 against 40.11). The chart score depends on a setting: Mistral OCR 4 scores 2.35 on charts without annotations and 40.11 with them.
Getting the right value into the right field. This is extraction, and it is what a downstream system consumes. A transcription score cannot tell you whether the invoice total landed in the total field, or whether row 1,800 of a long table survived. Mistral's published benchmarks measure transcription and layout, and its annotation step is a single pass with no field-level confidence. anyformat's long-document study held about 99% row recovery on tables running to 50 pages and roughly 2,400 rows (anyformat benchmark, 2026), and each extracted field carries a confidence and its source on the page.
If accuracy decides the choice, test it the way you will run it. Take 50 to 100 of your own documents, extract the same schema with both, and score at field level: exact match for identifiers and amounts, row recall for tables. Then check whether the confidence each system reports predicts which fields are wrong, because that decides how much a person has to review.
anyformat's evals suite runs that kind of comparison on your own documents before a workflow goes live.
Run both on your own documents. Try anyformat free, with a one-time grant of 50,000 credits on Pay As You Go, or book a demo.
Mistral publishes a flat per-page rate: $4 per 1,000 pages for OCR and $5 per 1,000 annotated pages (Mistral model card, 13 October 2026). There are no complexity tiers.
anyformat bills in credits, charged per page and per operator: 10 credits cost €0.01, so Parse is 25 credits per page (€0.025, published as $25 per 1,000 pages), Extract 35 credits per page, and Validate 5 credits per AI rule, once per extraction (anyformat pricing, 2026). Parse plus Extract is 60 credits, or €0.06, per page.
On raw OCR, Mistral is about six times cheaper per page, and that is a real difference. The two numbers are not the same product: Mistral's rate buys a model call, and anyformat's includes the review queue, validation and orchestration that you would otherwise build and run yourself.
Compliance and European sovereignty
Both vendors are European and both list ISO 27001. anyformat is ISO 27001 certified, GDPR-compliant and an EU-headquartered company with an EU legal entity, with compliance included at every plan and no surcharge. Mistral AI's trust center lists ISO 27001, ISO 27701 and SOC 2 Type II, with the certificates available there (Mistral Trust Center, 13 October 2026).
The detail to check is where processing runs. Mistral's trust center names its subprocessors and the regions they serve, including Microsoft, Google, CoreWeave and its own Mistral Compute, and lists a US API endpoint alongside the European ones. For retention terms and the route your contract uses, check Mistral's data processing terms directly; this page does not restate them.
Workflow builder and orchestration
This is the clearest gap. Mistral OCR is an endpoint, and orchestration is code your team writes and maintains. Classification, splitting multi-document packets, routing by document type, conditional logic, retries and human review are all engineering work.
anyformat includes Studio, a no-code workflow canvas: branching, conditions, splitting, routing, extraction operators, validation rules, cross-referencing against internal data sources, and human-in-the-loop review placed wherever the process needs it. Operations teams change the flow without a code deploy.
Splitting multi-document packets and routing each document by type are the kind of steps that otherwise become code.
Human review and validation
Mistral gives you bounding boxes and confidence scores as inputs; the review process itself is outside the product, and its documentation describes no review queue, field override or verification status.
anyformat ships the review loop: low-confidence fields route to a reviewer, who sees the value next to its source on the page and can correct it. Validation rules written in plain language run against the extracted values, with error and warning severity.
Version stability
mistral-ocr-latest is an alias that moves to new model versions as Mistral releases them. OCR 4.0 carries a deprecation date of 29 September 2026 with OCR 4.1 named as the replacement (Mistral model card, 5 October 2026). Mistral announces its retirements, which is the right way to do it, but an integration pinned to a retired model still has to be re-validated.
anyformat workflow versions are pinned and immutable: a published version_id does not change underneath you, which matters when a regulated process has to reprocess a document identically years later.
Yes, if you want extraction to run as a process rather than a call: schema-validated output with per-field confidence, a review queue, validation rules, no-code orchestration and version-pinned workflows. If you want markdown or text out of documents at volume and at the lowest price, it depends on the job: Mistral OCR does that well, and anyformat is not a cheaper way to do it. And if you like Mistral's parsing, the two can sit together, because the layer that makes the output production-safe is the part anyformat adds.
When to choose Mistral OCR
High-volume OCR or markdown conversion at the lowest price per page, with the pipeline around it as code your team is happy to own, or where your documents are in low-resource languages and transcription is the whole job.
When extraction has to run as a process rather than a call: per-field confidence with evidence, validation rules, human review, no-code orchestration and pinned workflow versions. Get started at anyformat.ai.
See also the other AI-native comparisons: anyformat vs LlamaParse and anyformat vs Reducto, or browse all comparisons.
