The headline price is a tie. AWS Textract Analyze Expense and the Azure prebuilt invoice model both cost $10 per 1,000 pages, and plain text OCR costs $1.50 per 1,000 on both clouds. So price does not decide this one. What decides it is language coverage, where Azure reads invoices in 27 languages against English only on Textract, the free tier, where Azure gives 500 pages every month forever and AWS gives 100 pages a month for three months, and accuracy, where the one independent invoice benchmark we could find puts Azure ahead at 93 percent field accuracy against 78 percent. Textract wins on published pricing you can actually read and on a 3,000 page ceiling per PDF. Neither one hands you a finished spreadsheet.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
Both clouds charge $10 per 1,000 pages to pull fields off an invoice. Teams read that, call it a wash, and pick on brand loyalty. The bill you actually receive at the end of the month diverges for reasons that never appear on either price list.
The official Azure pricing page renders every rate as a placeholder dollar sign with no figure attached, and tells you to sign in to the calculator instead. You cannot compare against a number you are not shown, and the figure changes with your enrollment. AWS publishes a full public rate card you can read in ten seconds.
Analyze Expense extracts invoice and receipt fields in English. General text detection covers six languages, and handwriting recognition is English only. Azure prebuilt invoice covers 27 languages. If any share of your vendors bill in another language, the price tie stops being a tie.
Textract synchronous calls accept a single page of PDF or TIFF at 10 MB. Multi page invoices force you onto the asynchronous path with S3 staging, job polling and completion notifications. That is real engineering the per page rate does not mention.
AWS gives Analyze Expense 100 pages a month and only for the first three months. Azure gives 500 pages a month with no expiry, though the free F0 tier reads just the first two pages of any document and caps files at 4 MB. Neither is enough to run on, but they support very different amounts of evaluation.
Textract returns SummaryFields and LineItemGroups. Azure returns a field collection with line items. Both are structured, and neither is a file your accounting system will accept. Vendor normalization, totals validation, duplicate checks and the export path are yours to build in either column.
Every figure below comes from the official AWS rate card, Microsoft documentation, or a named third party benchmark, with the assumptions written down so you can substitute your own volumes.
Analyze Expense is $10 per 1,000 pages and the Azure prebuilt invoice model is $10 per 1,000 pages. Plain OCR is $1.50 per 1,000 on both. Anyone telling you one cloud is dramatically cheaper for invoice extraction is comparing different products.
Azure prebuilt invoice reads 27 languages. Textract Analyze Expense reads English. For a US business paying only domestic vendors this changes nothing. For anyone with cross border suppliers it decides the whole question.
The BusinesswareTech invoice benchmark scored Azure at 93 percent field accuracy and 87 percent on line items, against 78 percent and 82 percent for Textract. Published January 2025 on scanned invoices, so treat it as directional rather than final.
Textract handles 3,000 pages per PDF asynchronously at up to 500 MB. Azure caps at 2,000 pages per document on the paid tier. Both are far beyond a normal invoice, and both matter if you batch statements or contracts through the same pipeline.
AWS lists every operation, every volume tier and every free tier allowance publicly. Azure requires you to sign in. For procurement approval and for budgeting a project before you have an account, that difference is not cosmetic.
Field mapping into your chart of accounts, line item splitting across continuation pages, totals validation, vendor matching and an xlsx, csv or QBO export. Identical backlog on both clouds, and usually larger than the API bill itself.
Public benchmarks describe someone else documents. Four steps get you an answer about yours.
Not the clean digital PDFs. The faxed one, the scan with a staple shadow, the vendor whose total sits inside a footer image, the three page one where line items continue across a break. These decide the outcome, and clean documents tell you nothing.
Tip: Include at least one multi page invoice. Continuation tables are where both services quietly lose rows.
If any invoice in the sample is not in English, Textract Analyze Expense is already out for that document and you can stop comparing. Run the same check on handwritten annotations, which Textract reads in English only.
Tip: This single question resolves the decision for a lot of teams before any accuracy testing happens.
Do not eyeball the JSON. Compare the extracted line item count and the extracted total against the real document. A service that transcribes every word but drops the fourth row of a table is worse for accounts payable than one that reads slightly less cleanly and keeps every row.
Tip: Row loss is the failure mode that costs money later, because a short invoice still reconciles until someone checks.
At $10 per 1,000 pages, 100,000 invoice pages a year is $1,000 on either cloud. Then add S3 staging and async job handling on AWS, or the equivalent on Azure, plus schema mapping, validation, retries and the export path. Estimate that in engineer weeks and add maintenance as vendor layouts drift.
Tip: For anything under a few hundred thousand pages a year, this line dwarfs the API bill.
Pricing, accuracy, capabilities and hard limits, from official documentation and independent benchmarks, with every source named.
Working out the true build cost on each cloud, including the async plumbing and the field mapping neither rate card mentions.
Asked whether an in house extraction build on the cloud you already pay for beats a flat rate tool, and needing a number that survives review.
Checking whether staying inside an existing AWS or Azure agreement costs accuracy, and by how much, before raising it as a question.
Needing invoice extraction beyond English, where the two services are not remotely equivalent regardless of price.
Both cost $10 per 1,000 pages to extract invoice fields, so price is a tie. Azure Document Intelligence wins on language coverage (27 languages against English only), on the free tier (500 pages a month with no expiry against 100 pages a month for three months), and on the one independent invoice benchmark we could find (93 percent field accuracy against 78 percent). AWS Textract wins on published, readable pricing, on a 3,000 page ceiling per PDF, and on a 40 plus field normalized schema.
The practical rule: if every vendor invoices you in English and you are already deep in AWS, Textract is a defensible choice and the price is identical. If any meaningful share of your invoices arrive in another language, Azure is not just better, it is the only one of the two that works. Neither service returns a finished spreadsheet, and that gap costs the same on both clouds.
AWS publishes a full public rate card. Microsoft does not, so the Azure figures below are the rates in general circulation for the S0 tier and should be confirmed in the Azure pricing calculator against your own enrollment before you budget on them. AWS figures are US West (Oregon), first 1,000,000 pages.
| Task | AWS Textract | Azure Document Intelligence |
|---|---|---|
| Plain text OCR | $1.50 (Detect Document Text) | $1.50 (Read) |
| Invoice fields and line items | $10.00 (Analyze Expense) | $10.00 (prebuilt invoice) |
| Tables | $15.00 ($0.015 per page) | Included in Layout |
| Forms and key value pairs | $50.00 ($0.05 per page) | Included in prebuilt models |
| Custom model analysis | $25.00 (Custom Queries) | $30.00 (custom extraction) |
| Volume tier past 1M pages | $8.00 Analyze Expense, $0.60 plain OCR | Commitment tiers, not published |
| Free tier | 100 pages a month, first 3 months | 500 pages a month, ongoing |
AWS rates come from the official Amazon Textract pricing page checked in August 2026. Two details matter beyond the headline. Textract charges Layout at no cost when you combine it with Tables, which is unusual and genuinely useful. And the free tier difference is larger than it looks: Azure lets you keep evaluating at 500 pages a month indefinitely, while the AWS Analyze Expense allowance is 100 pages a month and expires after three months.
A note on the Azure column. The official Microsoft pricing page for Document Intelligence currently displays every rate as a placeholder with no dollar figure and directs you to sign in to the calculator. We publish $10 per 1,000 pages for prebuilt invoice because that is the figure consistently reported and it matches the Layout and prebuilt tier, but we cannot verify it against a public Microsoft page, and we would rather say so than pretend otherwise.
Vendor self reported accuracy numbers are worth very little, because each vendor picks the dataset. The most useful independent comparison we found is the BusinesswareTech invoice extraction benchmark, published January 2025, which ran scanned invoices in JPG, PNG and PDF without text layers, spanning documents from 1971 to 2020, across at least three layout types, and scored output against human verified ground truth.
| Metric | AWS Textract (Analyze Expense) | Azure Document Intelligence |
|---|---|---|
| Field level accuracy | 78% | 93% |
| Line item detection | 82% | 87% |
Read that with the caveats attached. The benchmark is from January 2025 and both services have shipped model updates since. The exact number of invoices tested is not disclosed. And the dataset is deliberately hard, built from scanned documents with no text layer, which is the scenario where Textract accuracy is known to fall furthest. On clean digital PDFs the gap narrows considerably. What the benchmark does support is a directional claim: on difficult scanned invoices, Azure extracted more fields correctly out of the box and needed less post processing.
This is where the two services stop looking similar. Everything below is from AWS and Microsoft documentation.
| Capability | AWS Textract | Azure Document Intelligence |
|---|---|---|
| Invoice model | Analyze Expense | prebuilt-invoice, v4.0 GA |
| Line items returned natively | Yes, LineItemGroups | Yes |
| Normalized fields | 40 plus | Prebuilt invoice field set |
| Invoice languages | English | 27 languages |
| General text languages | English, French, German, Italian, Portuguese, Spanish | Broader, varies by model |
| Handwriting | English only | Supported |
| Max pages per document | 3,000 (async) | 2,000 (paid S0) |
| Max file size | 500 MB async, 10 MB sync | 500 MB paid, 4 MB free |
| Synchronous PDF limit | 1 page | Not applicable |
| Free tier page cap per document | Full document | First 2 pages only on F0 |
| Input formats | PDF, TIFF, JPEG, PNG | PDF and images for prebuilt |
| Vertical text | Not supported | Not applicable |
| Public price list | Yes, full rate card | No, placeholder values only |
Two limits catch teams out. On AWS, the synchronous API accepts a single PDF page, so any multi page invoice pushes you to the asynchronous flow with S3 staging and job polling. On Azure, the free F0 tier reads only the first two pages of any document, which means a free tier evaluation on three page invoices silently tests less than you think it does.
For invoices specifically, Azure Document Intelligence is the stronger default. It reads 27 languages against English only, it scored higher on the one independent benchmark available, and its free tier lets you evaluate properly. AWS Textract is the better pick when you are already standardized on AWS, all your invoices are English, and you value a public rate card you can budget from without an account.
No, they cost the same for invoice extraction. Both charge $10 per 1,000 pages for their invoice model and $1.50 per 1,000 pages for plain text OCR. AWS is cheaper at very high volume, dropping to $8 per 1,000 pages past one million pages, but Microsoft offers commitment tiers that are not published. Treat the price as a tie and decide on capability.
Azure AI Document Intelligence is the direct equivalent, and it was called Azure Form Recognizer until Microsoft renamed it. The mapping is close: Textract Detect Document Text corresponds to the Azure Read model, Textract Analyze Expense corresponds to the Azure prebuilt invoice and receipt models, and Textract Analyze Document with Forms and Tables corresponds to Azure Layout.
Yes. Form Recognizer is the former name for what Microsoft now calls Azure AI Document Intelligence. The service, the models and the API are the same product line, and the current invoice model is prebuilt-invoice at version 4.0, which reached general availability on 30 November 2024. Older tutorials and Stack Overflow answers still use the Form Recognizer name.
Yes. The Analyze Expense API returns LineItemGroups containing individual line items alongside SummaryFields for header level data such as vendor name, invoice date and total. It normalizes more than 40 field types, including ITEM, QUANTITY and PRICE within line items, and puts any unmatched text such as a SKU or long description into an EXPENSE_ROW entry.
Plain text extraction with Detect Document Text costs $1.50 per 1,000 pages, and invoice extraction with Analyze Expense costs $10.00 per 1,000 pages, both in US West Oregon for the first million pages a month. Past one million pages those rates drop to $0.60 and $8.00 respectively. Forms cost $0.05 per page and Tables $0.015 per page.
On the BusinesswareTech benchmark of scanned invoices published in January 2025, Textract Analyze Expense scored 78 percent field level accuracy and 82 percent on line item detection. Accuracy is considerably higher on clean digital PDFs than on the low quality scans that benchmark focused on, so treat 78 percent as a floor for difficult documents rather than a typical result.
Yes, with different plumbing. Textract asynchronous operations accept PDF and TIFF up to 3,000 pages and 500 MB, but the synchronous API is limited to one page, so multi page documents require the async flow with S3. Azure handles up to 2,000 pages per document on the paid tier, though the free F0 tier processes only the first two pages.
Both of these are extraction layers, not invoice products. You are buying structured JSON, and then building the thing your finance team actually asked for on top of it: mapping fields into your chart of accounts, splitting line items that continue across a page break, validating that the rows sum to the stated total, matching vendors against records you already hold, and writing a file your accounting system will import. That is a reasonable trade when you have engineers, volume and a reason to keep the data inside your own cloud account.
It is a poor trade when what you needed was invoice data in a spreadsheet by Thursday. InvoiceExtractor runs the extraction, the field mapping, the line item splitting and the totals validation, then exports xlsx, csv, json or QBO. Starter is $49 a month for 2,500 Base AI pages, which works out around $19.60 per 1,000 pages, and Plus is $149 a month for 10,000 pages, around $14.90 per 1,000.
That is more per page than either cloud API. It is considerably less than either cloud API plus the engineer who turns raw JSON into something your ledger accepts, plus the ongoing maintenance as vendor layouts change. Which side of that line you sit on depends entirely on whether extraction is your product or your chore.
For invoices, Azure Document Intelligence is the stronger default. It reads 27 languages against English only on Textract Analyze Expense, it scored 93 percent field accuracy against 78 percent on the independent BusinesswareTech benchmark, and its free tier allows 500 pages a month with no expiry. AWS Textract is the better pick if you are already standardized on AWS, every invoice is in English, and you want a public rate card you can budget from.
No, they cost the same for invoice extraction. Both charge $10 per 1,000 pages for their invoice model and $1.50 per 1,000 pages for plain text OCR. AWS drops to $8 per 1,000 pages past one million pages a month, and Microsoft offers commitment tiers that it does not publish. Treat the headline price as a tie and decide on capability instead.
Azure AI Document Intelligence, formerly called Azure Form Recognizer. The models map closely: Textract Detect Document Text corresponds to the Azure Read model, Textract Analyze Expense corresponds to the Azure prebuilt invoice and receipt models, and Textract Analyze Document with Forms and Tables corresponds to Azure Layout.
Yes. Form Recognizer is the former name for Azure AI Document Intelligence. It is the same product line, and the current invoice model is prebuilt-invoice version 4.0, generally available since 30 November 2024. Older tutorials and forum answers still use the Form Recognizer name, which is why the two names appear interchangeably in search results.
Yes. The Analyze Expense API returns LineItemGroups containing individual line items, alongside SummaryFields for header data such as vendor name, invoice date and total. It normalizes more than 40 field types including ITEM, QUANTITY and PRICE, and places unmatched text such as a SKU into an EXPENSE_ROW entry.
Detect Document Text costs $1.50 per 1,000 pages and Analyze Expense costs $10.00 per 1,000 pages, both in US West Oregon for the first million pages a month. Past one million pages those rates fall to $0.60 and $8.00. Forms cost $0.05 per page, Tables $0.015 per page, and Layout is free when combined with Tables.
On the BusinesswareTech benchmark of scanned invoices published January 2025, Textract Analyze Expense scored 78 percent field level accuracy and 82 percent on line item detection, against 93 percent and 87 percent for Azure. That dataset was built from low quality scans without text layers, so treat 78 percent as a floor for difficult documents rather than a typical result on clean digital PDFs.
Yes, with different plumbing. Textract asynchronous operations accept PDF and TIFF up to 3,000 pages and 500 MB, but the synchronous API handles only one page, so multi page invoices require the async flow with S3 staging and job polling. Azure processes up to 2,000 pages per document on the paid S0 tier, while the free F0 tier reads only the first two pages.
Every Analyze Expense and Detect Document Text rate, plus the volume tiers and free tier detail.
What the Azure prebuilt invoice model costs, and why the official page shows no numbers.
The other head to head: token billing against a flat per page rate, with independent accuracy scores.
Mistral OCR is $4 per 1,000 pages flat, well under both cloud invoice models.
Prefer an API that already returns mapped invoice fields? Skip the schema build entirely.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
billed as
$288 yearly
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
billed as
$888 yearly
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
billed as
$ yearly
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |
