Mistral OCR 4.1 charges a flat $4 per 1,000 pages, or $2 through the Batch API. AWS Textract charges $10 per 1,000 pages for Analyze Expense, the API built for invoices and receipts. Mistral is cheaper and returns markdown you still have to parse. Textract costs more and returns typed invoice fields, vendor, totals and line item groups, without you defining a schema. The expensive mistake is calling Textract the general way: AnalyzeDocument with Forms, Queries and Tables bills $70 per 1,000 pages for the same invoice, seven times the purpose built rate. Neither one hands you a finished spreadsheet.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
Both vendors quote dollars per 1,000 pages, which makes this look like an easy comparison. It is not, because the two numbers buy different amounts of finished work and the Textract rate card has sixteen entries.
There is no single Textract price. Detect Document Text is $1.50 per 1,000 pages, Analyze Expense is $10, Forms is $50, and Forms with Queries and Tables is $70. The number you pay depends entirely on which FeatureTypes you pass in the request, and most teams pick the wrong one first.
Reaching for AnalyzeDocument with FORMS and TABLES is the intuitive way to pull an invoice apart, and it is the most expensive path on the list. Analyze Expense returns more invoice structure for $10 per 1,000 pages. The general purpose bundle returns less structure for $70.
The $4 per 1,000 pages OCR endpoint gives you the page as markdown. Typed invoice fields come from the Annotations feature, billed as Document AI at $5 per 1,000 pages, and you write the schema yourself. Textract Analyze Expense ships an invoice schema already defined.
Mistral halves the rate to $2 per 1,000 pages through the Batch API. On AWS the synchronous and asynchronous meters are priced identically at every feature level, so running Textract as an overnight job saves nothing at all.
Textract drops Analyze Expense from $10 to $8 per 1,000 pages only above one million pages a month. Almost no accounts payable team reaches that, so the headline rate is the rate you pay for years.
Markdown and a JSON blob are not something your accounting system can import. Totals validation, line item splitting, vendor name normalization and the export to xlsx, csv or QBO is work you own in both columns, and it appears on neither price list.
Every AWS figure below comes from the AWS Price List API for us-east-1, document version 20260831092230 published 31 August 2026, because the public Textract pricing page no longer renders its own numbers. Mistral figures come from the published rate card.
The complete Textract rate card pulled from the AWS Price List API, first tier and above one million pages, so you can see exactly where Analyze Expense sits against the Forms and Tables bundles that cost up to eight times more.
Mistral gives you one number that does not move. Textract gives you a number that depends on the FeatureTypes array in your request. Both facts matter, and they matter to different people on the same team.
Analyze Expense returns SummaryFields and LineItemGroups typed for invoices and receipts. Mistral OCR returns markdown with bounding boxes. Document AI returns fields against a schema you supply. The output difference is the real story here.
Mistral Batch halves the bill to $2 per 1,000 pages. Textract has no batch discount and its volume break needs a million pages a month. At 10,000 invoices a month that is $20 against $100.
Textract runs in your chosen AWS region under your existing account, IAM and BAA. Mistral is a hosted European API. For a US healthcare or government buyer that difference often settles the choice before price is discussed.
No independent benchmark scores Mistral OCR and Textract head to head on US invoices. Rather than invent one, the page below sets out what each vendor and each third party test actually measured, and where the gaps are.
Published benchmarks describe someone else documents. Four steps get you a number for yours.
Not the clean digital PDFs. The faxed one, the scan with a staple shadow, the vendor whose totals sit in a footer image, the three page one where line items continue across pages. These decide the result.
Tip: Include at least one multi page invoice. Continuation tables are where extraction quietly loses rows.
Use the AnalyzeExpense operation rather than AnalyzeDocument with FORMS and TABLES. It is built for invoices, returns typed SummaryFields and LineItemGroups, and costs $10 per 1,000 pages instead of $70 for the general bundle.
Tip: If you have already built on AnalyzeDocument, this one change can cut the bill by seven times.
Do not eyeball the markdown. Count extracted line item rows and check totals against the real document. A model that reads every word but drops the fourth row of a table is worse for accounts payable than one that reads slightly less cleanly.
Tip: Count rows first, then totals. Row loss is the failure mode that costs money downstream.
Whichever wins, you still need schema mapping, validation, retries and an export path into your accounting system. Estimate that in engineer weeks and add maintenance as vendor layouts drift.
Tip: Below a few hundred thousand pages a year this line is almost always bigger than the API bill.
Pricing, output, accuracy, compliance and limits, from official rate cards and documentation, with every source named.
Working out which Textract FeatureTypes to call, what Mistral actually returns, and how much code sits between either output and a posting file.
Asked whether an in house extraction build beats a flat rate tool, and needing a per invoice number that survives review.
Weighing whether staying inside one cloud account, region and IAM boundary is worth paying two and a half times the Mistral rate.
Where data residency, an existing BAA and a named region decide the vendor before anyone opens the rate card.
Last updated September 2026
Mistral OCR 4.1 costs a flat $4 per 1,000 pages, or $2 through the Batch API. AWS Textract charges $10 per 1,000 pages for Analyze Expense, the operation purpose built for invoices and receipts. Mistral is two and a half times cheaper and returns the page as markdown that you still have to parse into fields. Textract costs more and returns typed invoice structure, vendor, dates, totals and line item groups, with no schema work from you. The trap that decides most real bills is not the vendor choice at all: calling Textract with AnalyzeDocument using Forms, Queries and Tables costs $70 per 1,000 pages for the same invoice, seven times the purpose built rate.
Put simply, Mistral wins on unit price and portability, Textract wins on invoice specific output and on running inside an AWS account you already trust. Neither one produces a file your accounting system can import without further work.
Both vendors quote per 1,000 pages, so this table needs no conversion. The AWS figures come from the AWS Price List API for us-east-1 rather than the public pricing page, which no longer renders its own numbers and returns an unresolved template placeholder instead.
| Option | Per 1,000 pages | What you get back |
|---|---|---|
| Mistral OCR 4.1 via Batch API | $2.00 | Markdown, bounding boxes, no typed fields |
| Mistral Libraries tool OCR | $3.00 | Markdown inside the Libraries product |
| Mistral OCR 4.1 standard API | $4.00 | Markdown, bounding boxes, no typed fields |
| Mistral Document AI (Annotations) | $5.00 | Typed fields against a schema you define |
| AWS Textract Analyze Expense | $10.00 | Invoice and receipt schema, line item groups |
| AWS Textract Forms | $50.00 | Generic key value pairs, no invoice typing |
| AWS Textract Forms, Queries and Tables | $70.00 | Generic pairs plus table cells, no invoice typing |
At 10,000 invoice pages a month, the spread runs from $20 on Mistral Batch to $100 on Textract Analyze Expense to $700 if you call Textract the general purpose way. That last column is not a hypothetical. It is the default result of reaching for the API that sounds like it does what you want.
Because Textract is not one price, it is sixteen. The rate is set by the FeatureTypes you pass in the request, and the invoice specific operation is far cheaper than the general document one. AnalyzeExpense is a separate API call built for invoices and receipts; AnalyzeDocument with FORMS and TABLES is a general extractor that costs seven times more and returns less invoice structure. Here is the full card.
| Textract feature | First 1M pages | Above 1M |
|---|---|---|
| Detect Document Text | $1.50 | $0.60 |
| Signatures | $3.50 | $1.40 |
| Layout | $4.00 | $3.00 |
| Analyze Expense (invoices) | $10.00 | $8.00 |
| Tables | $15.00 | $10.00 |
| Queries | $15.00 | $10.00 |
| Queries and Tables | $20.00 | $15.00 |
| Custom Queries | $25.00 | $15.00 |
| Analyze ID | $25.00 | $10.00 |
| Custom Queries and Tables | $30.00 | $20.00 |
| Forms | $50.00 | $40.00 |
| Forms and Queries | $55.00 | $45.00 |
| Forms and Custom Queries | $65.00 | $50.00 |
| Forms, Queries and Tables | $70.00 | $55.00 |
| Analyze Lending | $70.00 | $55.00 |
| Forms, Custom Queries and Tables | $80.00 | $60.00 |
Two details in that table are worth more than the headline. The synchronous and asynchronous meters are priced identically at every single feature level, so there is no saving in running Textract as an overnight batch. And the volume break needs a million pages a month, which almost no accounts payable department reaches, so the first tier rate is the rate you will pay indefinitely.
This is where the price gap earns or loses its keep. Textract Analyze Expense returns an invoice typed structure without you describing it. Mistral OCR returns the page as markdown with spatial data, and typed fields require the separate Document AI product plus a schema you write.
| Capability | Mistral OCR 4.1 | AWS Textract Analyze Expense |
|---|---|---|
| Native invoice field typing | No, markdown only | Yes, SummaryFields |
| Line item groups | Present in markdown, not typed | Yes, LineItemGroups |
| Bounding boxes | Yes | Yes, Geometry blocks |
| Confidence scores | Yes, per block | Yes, per field |
| Schema you must define | Yes, for Document AI | No |
| Batch discount | Yes, 50 percent | None |
| Language coverage | 170 languages | English plus 7 more on Expense |
| Runs in your own cloud account | No, hosted API | Yes, your AWS region |
| Finished spreadsheet or QBO file | No | No |
Read that last row twice. Both products stop at structured output. Turning that into something an accountant can post means validating that line items sum to the invoice total, normalizing vendor names so the same supplier does not appear three ways, handling credit notes, and writing the export. That backlog is identical whichever column you pick, which is the honest argument for skipping the model layer entirely.
No independent benchmark scores these two head to head on US invoices, and any page that gives you a single percentage for both is making it up. What does exist is partial and worth knowing. In BusinesswareTech testing on scanned documents, AWS Textract reached 78 percent on header fields and 82 percent on line items, which is a notably strong line item result, beaten there only by Azure Document Intelligence at 87 percent. On OCR Arena community voting, Mistral OCR v3 sits at ELO 1576 and rank nine, behind the frontier vision models. Those two measurements use different documents, different versions and different scoring, so they cannot be subtracted from one another.
The practical read: Textract is comparatively strong at line items, which is the hard part of an invoice and the part that costs you money when it fails. Mistral is strong at general document reading across 170 languages. If your invoices are US English and line item accuracy is the thing you are buying, that leans toward Textract despite the price. If you process multilingual documents or need to self host eventually, that leans the other way.
Textract runs inside your AWS account, in a region you pick, under your existing IAM policies and, where applicable, your existing AWS Business Associate Addendum. For a US healthcare provider, a bank or a government contractor, that is often the entire decision, made before anyone compares rates. Mistral is a hosted European API, which suits teams with European data residency requirements and complicates life for US buyers with strict domestic residency clauses.
There is a second practical difference. Adopting Textract means an AWS account, S3 buckets for documents, IAM roles, region selection and the glue code around all of it. Adopting Mistral means an API key. For a small finance team without engineers, that gap is larger than the price difference. Teams whose document pipeline also has to handle other financial paperwork usually discover the same pattern elsewhere: converting a PDF bank statement into a spreadsheet is a genuinely different extraction job with its own tooling, and trying to make one general OCR call serve every document type is how these projects sprawl.
Choose AWS Textract Analyze Expense if you are already on AWS, your invoices are US English, line item accuracy matters more than unit cost, and you need documents to stay inside your own cloud account and region. Budget $10 per 1,000 pages and make sure your code calls AnalyzeExpense rather than AnalyzeDocument.
Choose Mistral OCR if unit cost dominates, you process multiple languages, your workload can run as an overnight batch at $2 per 1,000 pages, and you are comfortable writing the schema and field mapping yourself.
Choose neither if what you actually want is invoices arriving as clean rows in a spreadsheet or your accounting system. Both of these are components, not products. They return structured data to a developer, and every hour after that point is yours to spend. The tool at the top of this page does the whole job, including line item splitting, totals validation and export, at a flat monthly rate with no AWS account, no schema definition and no per page meter.
Mistral, by two and a half times on the standard rate and five times on batch. Mistral OCR 4.1 is $4 per 1,000 pages, or $2 through the Batch API. AWS Textract Analyze Expense is $10 per 1,000 pages for the first million pages a month. The gap widens sharply if you call Textract with Forms and Tables instead, which costs $70 per 1,000 pages.
Analyze Expense costs $10.00 per 1,000 pages for the first one million pages a month, dropping to $8.00 per 1,000 above that. Synchronous and asynchronous calls are priced identically, so there is no batch saving. The figures come from the AWS Price List API for us-east-1, because the public Textract pricing page no longer renders its own rates.
It reads invoices well, but it does not return invoice fields. Mistral OCR 4.1 gives you the page as markdown with bounding boxes and confidence scores. To get typed values like vendor, invoice number and total you use the Document AI Annotations feature at $5 per 1,000 pages and define the schema yourself. Textract Analyze Expense ships that schema already built.
No. The asynchronous meters are priced exactly the same as the synchronous ones at every feature level, verified from the AWS Price List API. Running Textract as an overnight job saves nothing. Mistral does halve its rate through the Batch API, from $4 to $2 per 1,000 pages, which is one of the larger practical differences between the two.
No. Despite Mistral publishing open weight language models, Mistral OCR is a hosted commercial API only. There are no downloadable OCR weights, so self hosting it on your own hardware is not currently possible. If running extraction inside your own infrastructure is a requirement, that rules it out.
Partly. Textract supports handwriting in English through Detect Document Text and Forms, and handwriting recognition is generally weaker than print across every vendor. For invoices, handwritten amounts and dates should be treated as requiring human review regardless of which API reads them, because a misread digit in a total is an expensive error to catch late.
Mistral OCR covers 170 languages, which is its clearest advantage for international document sets. AWS Textract Analyze Expense supports English plus a smaller set of additional languages, with broader coverage on the plain text detection API than on the invoice specific one. For US only accounts payable this difference rarely matters.
No. Both return structured data to a developer, not a file an accountant can post. You still need to validate that line items sum to the invoice total, normalize vendor names, handle credit notes, and write the export to xlsx, csv or QBO. That work is the same size whichever of the two APIs you choose.
Mistral OCR is $4 per 1,000 pages flat, $2 on Batch. The full breakdown and what the rate excludes.
The complete sixteen line Textract rate card, including the Forms and Tables trap on invoices.
What to use instead of Textract when you want finished invoice rows rather than JSON.
The other side of the Textract comparison, against the Azure prebuilt invoice model.
How Mistral prices against Gemini, the cheapest credible option on a per page basis.
Prefer an API that already returns invoice fields? Skip the schema build entirely.
Read PDFs, scans and photos with AI based invoice OCR, without writing the pipeline.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
per month
$288 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 2,500 pages |
| Pro AI Best accuracy | 500 pages |
Scale invoice extraction across your whole team with automation.
USD
per month
per month
$888 charged today
Choose speed vs accuracy when extracting
| Base AI Faster | 10,000 pages |
| Pro AI Best accuracy | 2,000 pages |
Enterprise‑grade invoice extraction, security, and controls.
USD
per month
per month
$ charged today
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |
Secure checkout powered by Stripe
