Research-backed calculator guide

Mistral OCR Pages vs Annotations: Budget the Real Document Pipeline

Estimate Mistral OCR and Document AI spend per 1,000 pages, then add retries, downstream parsing, storage, and human review for a more realistic document-pipeline budget.

What is billed: pages first, annotations second

Mistral’s pricing surface mixes several billing units. Its public pricing page says most models are priced per million input and output tokens, but OCR is priced per 1,000 pages, speech models per minute, and tool APIs per call. It also says batch processing reduces prices by 50% for high-volume work, with the exact eligibility and route worth confirming in the current Studio account. ^[1] Do not convert an OCR page price into tokens just because the response contains Markdown or JSON.

The first line in your estimate is therefore:

OCR page cost = billable pages ÷ 1,000 × published OCR rate

The page count should represent what the provider bills, not merely the number of files in your queue. A 200-file batch may contain 20,000 pages. Conversely, a one-page image and a 30-page PDF are different page quantities even though each is one API request. Count pages after normalizing your input inventory, and define how you handle blank pages, failed uploads, duplicate submissions, and retries.

The second decision is the output layer. Basic OCR extracts text and structured content from PDFs and images. The processor can return Markdown, tables, hyperlinks, image information, page metadata, blocks, and confidence scores. ^[4] OCR 4 or newer supports block extraction with labels such as text, title, list, table, image, equation, caption, code, references, header, footer, and signature. That structure can reduce downstream parsing work, but it does not turn every page into a token-priced text completion.

Annotations add a schema-driven interpretation layer. Mistral documents two modes. bbox_annotation annotates bounding boxes extracted by OCR, such as charts and figures, according to a user-provided format. document_annotation returns an annotation for the whole document in a supplied format. The documentation says document annotation sends OCR Markdown plus the first eight extracted image bounding boxes to a vision-capable language model, while bounding-box annotation calls a vision-capable model for the relevant boxes after regular OCR. ^[3]

That distinction matters for budgeting. A page with many figures may invoke more image interpretation work than a text-heavy page. A document-level schema may create more structured output and validation effort than raw OCR. Yet the public documentation does not give a general table such as “$X per bbox” or “$Y per annotation field.” Use the published Document AI page rate where it applies, and separately model your own downstream validation and failure costs. Do not invent a per-box meter.

Current price references and version discipline

The current API pricing page is the cleanest place to start a version-aware estimate. It shows OCR 4.1 at $4 per 1,000 pages and OCR 3 at $2 per 1,000 pages, with no output-token price shown for those OCR rows. ^[2] Those are materially different inputs to a forecast. If your account or integration still targets OCR 3, use the OCR 3 row and label that assumption. If it follows the latest alias, confirm which model the account resolves to before comparing a historical invoice with a new forecast.

The OCR 4 launch article is useful for the structured-processing distinction. It lists OCR 4 API processing at $4 per 1,000 pages, a 50% Batch API discount that reduces that OCR line to $2 per 1,000 pages, and Document AI at $5 per 1,000 pages. ^[5] The article also says OCR 4 exposes bounding boxes, block types, and confidence scores, while Document AI parameters add schema-shaped JSON, image annotation, or a custom prompt on the same endpoint. These statements describe product behavior and published launch pricing; they are not an invoice guarantee for every future model alias, region, plan, or reseller.

Older search results can make the comparison confusing. The OCR 3 launch article reports $2 per 1,000 pages for OCR 3, $1 per 1,000 pages through Batch API, and a displayed OCR 3 card showing $3 per 1,000 pages for annotations. That is useful historical evidence, but it should not overwrite the current OCR 4.1 row or be blended into an OCR 4.1 forecast. ^{[2] [5]} Record the model, price page, currency, and checked date in the estimate.

Worked cost example

Assume a US-based team processes 25,000 pages in a month. The table is an original budgeting worksheet, not a vendor quote. It keeps published prices separate from operating assumptions so that you can replace each input with your invoice data.

The $125 Document AI line comes from the published OCR 4 launch figure. The $35 downstream line is an explicit estimate, not a Mistral OCR quote: it assumes 2,000 tokens of downstream input per page and a $0.70 per-million-token rate. The review and storage lines are also assumptions. If the downstream step uses a different model, a self-hosted parser, or no language-model post-processing, replace or remove that line.

The estimated effective cost is $329.58 ÷ 25,000 = $0.01318 per submitted page, before taxes and any fixed subscription or platform charges. If you measure only successful, accepted pages and 2% fail quality review, the denominator changes to 24,500 pages, or about $0.01345 per accepted page. That is why “cost per 1,000 pages” and “cost per accepted document” answer different finance questions.

A batch scenario should be shown as a separate case, not silently netted into the baseline. If your specific OCR route is eligible for the cited 50% batch reduction, the OCR 4.1 baseline would be modeled at $2 per 1,000 pages, or $50 for 25,000 pages. The OCR 4 launch page explicitly describes that discount for OCR API processing. It does not establish that every Document AI configuration, account, or future alias will receive the same reduction, so keep a conservative $125 Document AI case until eligibility is verified. ^[5]

A practical calculator method

Start with four inventory fields: documents, pages per document, expected retries, and annotation mode. Convert files into billable pages before multiplying by a rate. Then make the model version a required field rather than a note hidden in a comment. This prevents a historical OCR 3 benchmark from being compared with an OCR 4.1 invoice.

Next, split the pipeline into cost buckets:

1. Provider page processing. Use the current published OCR or Document AI per-1,000-page figure and record whether it is standard or batch.
2. Failure and retry work. Apply a retry percentage to pages, not only requests, when a partial document can be reprocessed. Track whether your client retries after timeouts or after receiving a structurally invalid response.
3. Annotation and validation. Include schema validation, malformed JSON repair, image review, and queue handling. The public annotation docs explain what the modes do, but they do not define your acceptance rate or human-review time. ^[3]
4. Downstream parsing or indexing. Estimate tokens or compute time only for the step that actually consumes OCR output. Do not count OCR output tokens as though Mistral’s OCR row were token-metered.
5. Storage, transfer, and retention. Add the cost of original files, rendered images, extracted tables, JSON results, logs, and backups if they are retained.
6. People and quality. Measure review seconds per page or per field, then apply the hourly rate used by finance. A cheap page API can still produce an expensive accepted-record workflow if every invoice needs manual correction.

The model pricing comparison can help when the downstream parser uses another model family. If the final step is token-based, use the AI token cost calculator to separate input, output, and cached-token assumptions. For the broader monthly view, compare the result with the LLM cost calculator rather than forcing page-based OCR into a token-only worksheet.

What to monitor after launch

Log the model alias returned by each response, submitted page count, annotation mode, batch flag, retry reason, and accepted-page status. The OCR processor response includes usage information, and Mistral’s pricing page says Studio provides usage and spending tracking for services including OCR. ^[1] A monthly reconciliation should compare those application counters with the provider dashboard and invoice period.

Watch three ratios. First, retry pages ÷ submitted pages shows reliability leakage. Second, reviewed pages ÷ submitted pages shows the labor multiplier. Third, accepted pages ÷ submitted pages gives a denominator for unit economics that is more honest than API pages alone. If a document contains many images or tables, split those cohorts because bounding-box and schema behavior can differ from plain text pages.

The published price is also not the full TCO. Exclusions from the worked example include taxes, account credits, minimum commitments, reseller markups, self-hosting infrastructure, queue latency, failed upload bandwidth, encryption and retention controls, and any commercial support plan. The article’s USD figures are date-stamped to November 2, 2026; verify the current pricing page and your workspace settings before approving a budget.

FAQ

Is Mistral OCR priced per token or per page?

Mistral’s API pricing documentation lists OCR models by price per 1,000 pages, not by input and output token rows. ^[2] OCR responses contain text and structured data, but that response shape does not change the OCR billing unit. Downstream language-model processing may introduce a separate token-based line.

What is the difference between OCR and document annotations?

Basic OCR extracts text and document structure. bbox_annotation adds structured interpretation for OCR-extracted bounding boxes, while document_annotation returns structured JSON for the whole document based on a schema. ^[3] Choose raw OCR when your own parser can use Markdown, tables, blocks, and confidence data. Choose annotations when the provider’s schema-shaped output reduces your application-side extraction work.

Is there a separate price for every annotation or bounding box?

The official annotation documentation explains the processing modes but does not publish a universal per-bounding-box price. ^[3] The OCR 4 launch page lists Document AI at $5 per 1,000 pages and describes additional schema and image-annotation capabilities. ^[5] Use that published page rate where applicable, then add your own validation and review assumptions. Do not multiply a made-up bbox price by the number of figures.

How much does Mistral OCR cost for 1,000 pages?

At the checked date, the API pricing page lists OCR 4.1 at $4 per 1,000 pages and OCR 3 at $2 per 1,000 pages. ^[2] The OCR 4 launch page lists Document AI at $5 per 1,000 pages. ^[5] Your estimate must identify the model version and whether you need the Document AI layer. Historical launch prices are not interchangeable with today’s model row.

Does batch processing halve annotation cost?

Mistral states a 50% Batch API discount for OCR 4 API processing in its launch article, reducing the cited $4 OCR rate to $2 per 1,000 pages. ^[5] That statement is not enough to assume every Document AI request receives the same treatment. Mark batch eligibility as an assumption until the current account, endpoint, and workload confirm it.

What should I include besides the page rate?

Include retries, schema validation, downstream model or compute work, storage and transfer, and human review. For a production budget, also track accepted pages, not just submitted pages. The worked example above turns those inputs into an effective cost per accepted page so an inexpensive API line does not hide an expensive correction queue.