Every OCR and document extraction rate, normalized to one unit so you can actually compare them. Plain OCR is $1.50 per 1,000 pages at Azure, AWS and Google. Structured extraction runs $5 to $70. The per-document, per-block and per-token vendors are converted onto the same axis below.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
Free plan extracts the first 5, rest can be unlocked after
Uploading...
A rate per 1,000 pages cannot tell you whether the fields come out right on your documents. Drop one in and look at the output before you pick a vendor.
Plain OCR costs $1.50 per 1,000 pages. That is the published rate at Azure AI Document Intelligence (Read), AWS Textract (Detect Document Text) and Google Document AI (Enterprise Document OCR), identical to the cent, and it falls to $0.60 per 1,000 at high volume. Reading text is a commodity and price is not a reason to pick one cloud over another. The money is in structured extraction, and there the spread is enormous: a prebuilt model is $10 per 1,000 pages, custom extraction is $30 at Azure and Google, AWS Textract Forms is $50 and Forms plus Tables plus Queries is $70, while Mistral Document AI returns structured output at $5 per 1,000 pages, the cheapest published structured rate anywhere. The lowest verified rate from any major vendor is Azure Read on its 16-million-page commitment tier at $0.45 per 1,000 pages, a rate Microsoft does not show on its pricing page.
Whichever you pick, the rate buys an API response. Classification, review and export are still yours to build.
You cannot rank OCR vendors on price until you put them on the same unit, and most of them refuse to use it. Azure, AWS and Google bill per page. Veryfi bills per document. Nanonets bills per block run. Gemini bills per token. Google's own prebuilt parsers bill in 10-page blocks. Converting all of that to dollars per 1,000 pages is what this page does, and it reorders the leaderboard.
Google prebuilt Invoice, Expense and Utility parsers charge $0.10 per 10-page block. Batch 10 pages a request and that is $10 per 1,000 pages. Send a 1-page invoice as its own request and it consumes a whole block, which is $100 per 1,000 pages. Same list price, ten times the cost, decided entirely by how you batch.
A 1-page receipt at $0.08 normalizes to $80 per 1,000 pages, over fifty times Azure Read. A 15-page bank statement at $0.25 normalizes to $16.67 per 1,000 pages, cheaper than Azure Layout. The same vendor is the most expensive or one of the cheapest, depending only on how long your documents are.
Google documents 258 tokens per document page, so a per-page figure has to be computed from token rates. Output tokens are 92% to 96% of a real OCR bill, which is why guides quoting input-only numbers land about thirteen times too low. Assuming 750 output tokens a page, Gemini 2.5 Flash-Lite works out to roughly $0.33 per 1,000 pages.
Nanonets bills per block run, a workflow step, at $0.02 to $0.30. Its own documentation puts a typical invoice at 4 to 6 blocks. There is no page in that unit, so any per-1,000-pages figure for Nanonets is invented. We do not publish one.
Cloud rates re-verified July 13, 2026 from primary sources: Microsoft's official Azure Retail Prices API (Document Intelligence and the newer Content Understanding), the AWS pricing pages (Textract in US West Oregon, plus the newer Amazon Bedrock Data Automation) and the Google Document AI pricing page. Mistral, Gemini, Google Cloud Vision and Veryfi rates were verified from their own pricing pages in July 2026. Rates move, so confirm anything you are about to sign against the vendor's page.
Two entries in the table below reward a second look, because both are billing-unit traps rather than headline rates. Google's prebuilt Invoice Parser bills per 10-page block, so a one-page invoice consumes a whole block and 1,000 of them cost $100 rather than $10. And Amazon Bedrock Data Automation is the only service here that charges more for a wider schema, adding $0.0005 per field per page above 30 fields, which takes a maxed-out 100-field blueprint to $75.00 per 1,000 pages. Full detail on Google's Invoice Parser pricing and the Bedrock Data Automation rate reference.
The newest category, agentic document extraction, is the priciest lane on this table. LlamaParse, Reducto and Landing AI all run a language model over each page and all price in credits, so their real per-page cost is buried. Normalized, agentic parsing runs $12.50 to $60 per 1,000 pages, roughly 8 to 40 times legacy OCR. See the full breakdown on the agentic document extraction pricing comparison.
| Vendor | Model or meter | Published unit | Normalized per 1,000 pages | Note |
|---|---|---|---|---|
| Azure AI Content Understanding | Minimal (digital DOCX, XLSX, HTML, TXT) | Per page | $0.01 | No OCR performed, so none is billed |
| Self-hosted open model | olmOCR on rented GPUs, saturated | Per GPU-hour | About $0.18 (published) | Only at 100% utilization. Text, not fields |
| Self-hosted open model | olmOCR on AWS on-demand H100s, saturated | Per GPU-hour | About $0.45 (computed) | Only at 100% utilization. Text, not fields |
| Azure AI Content Understanding | Basic (OCR on a scan) | Per page | $1.00 | The cheapest cloud OCR rate published |
| Azure AI Content Understanding | Standard (layout) | Per page | $5.00 | Half the Document Intelligence Layout rate |
| Azure AI Document Intelligence | Read (OCR) | Per page | $1.50 | $0.60 above 1M pages a month |
| AWS Textract | Detect Document Text | Per page | $1.50 | $0.60 above 1M pages a month |
| Google Document AI | Enterprise Document OCR | Per page | $1.50 | $0.60 above 5M pages a month |
| Google Cloud Vision | Document Text Detection | Per image (a PDF page is one image) | $1.50 | $0.60 above 5M units a month |
| Mistral | OCR 4 | Per page | $4.00 | Batch 50% off |
| Mistral | Document AI (structured) | Per page | $5.00 | Cheapest published structured rate |
| Gemini 2.5 Flash-Lite | Vision tokens | Per token (258 tokens a page) | About $0.33 (computed) | Assumes 750 output tokens a page |
| GPT-5.4-nano | Image patches | Per token (1,496 patches a page) | About $1.67 (computed) | About $0.84 through the Batch API |
| GPT-5.4-mini | Image patches | Per token (1,496 patches a page) | About $5.19 (computed) | Output is 65% of the bill |
| Claude Haiku 4.5 | Vision tokens | Per token (about 1,560 a page) | About $5.31 (computed) | Standard tier downscales the page |
| Claude Sonnet 5 | Vision tokens | Per token (2,714 a page) | About $19.39 (computed) | About $12.93 on the intro rate |
| Claude Opus 4.8 | Vision tokens | Per token (2,714 a page) | About $32.32 (computed) | Output is 58% of the bill |
| GPT-5.5 / GPT-5.6-sol | Image patches | Per token (2,080 patches a page) | About $32.90 (computed) | No image multiplier published |
| Azure AI Document Intelligence | Layout | Per page | $10.00 | Tables and structure |
| Google Document AI | Layout Parser | Per page | $10.00 | No volume step-down |
| Self-hosted open model | One AWS L4 instance at 50,000 pages a month | Per GPU-hour | About $11.75 (computed) | The same GPU, mostly idle. This is the real rate |
| AWS Textract | Tables | Per page | $15.00 | $10.00 above 1M pages |
| AWS Textract | Queries | Per page | $15.00 | Single tier |
| Azure AI Document Intelligence | Prebuilt (invoice, receipt, W-2, 1099) | Per page | $10.00 | $7.50 on the 1M commitment tier |
| Google Document AI | Prebuilt Invoice / Expense / Utility | Per 10-page BLOCK ($0.10) | $10.00 batched, $100.00 unbatched | A 1-page invoice consumes a full block |
| AWS Textract | Analyze Expense | Per page | $10.00 | $8.00 above 1M pages |
| Amazon Bedrock Data Automation | Standard output (text, tables, summary) | Per page | $10.00 | 6.7x the Textract OCR rate. No volume tier |
| Google Document AI | Form Parser | Per page | $30.00 | $20.00 above 1M pages |
| Azure AI Document Intelligence | Custom extraction | Per page | $30.00 | $20.00 above 1M, $18.00 committed |
| Google Document AI | Custom Extractor | Per page | $30.00 | Plus $0.05/hour idle hosting |
| Amazon Bedrock Data Automation | Custom output (blueprint, up to 30 fields) | Per page | $40.00 | Priciest flat custom rate of the big three |
| AWS Textract | Forms | Per page | $50.00 | $40.00 above 1M pages |
| AWS Textract | Forms + Tables + Queries | Per page | $70.00 | $55.00 above 1M pages |
| Amazon Bedrock Data Automation | Custom output, 100-field blueprint (the max) | Per page | $75.00 | +$0.0005 a field above 30. No other vendor does this |
| Azure AI Document Intelligence | Document classifier | Per page | $3.00 | Cheapest classifier of the three |
| Google Document AI | Custom Classifier / Splitter | Per page | $5.00 | $3.00 above 1M pages |
| Veryfi | Receipt | Per document (up to 15 pages) | $80.00 on 1-page receipts | $0.08 a document |
| Veryfi | Bank statement / check | Per document (up to 15 pages) | $16.67 on 15-page statements | $0.25 a document, rewards long files |
| Nanonets | Workflow block run | Per block run, not per page | Not expressible per page | $0.02 to $0.30 a block, 4 to 6 blocks an invoice |
| LlamaParse | Agentic parse | Per page (1,000 credits = $1.25) | $12.50 | Cheapest true agentic parse. Fast tier is $1.25, text only |
| Reducto | Agentic parse | Per page ($0.015 a credit) | $30.00 to $60.00 | Standard to complex. Standard parse is $15.00 |
| Landing AI ADE | DPT-2 parse | Per page ($0.01 a credit) | $30.00 | The 0.5-credit figure people quote is Classify, not parse |
| Extend | Extract (Performance) | Per page ($0.0125 a credit) | $62.50 | Extract 3 + auto-parse 2 credits; Light mode is $15.00 |
| Unstructured | Partition, any strategy | Per page ($0.03 a page) | $30.00 | Flat across strategies; $3,000 cap makes it $3.00 at 1M pages |
| DocuOCR | Product plans | Per page (workflow included) | About $14.00 to $20.00 | Classification, review and export bundled |
Prices are US list rates in USD, excluding tax. Azure figures are East US; a few Azure regions differ slightly. AWS figures are US West (Oregon).
The self-hosted rows deserve a word of warning, because they are the only ones in this table that are not really a price. An open model bills in GPU hours, so its per-page rate is the monthly hardware bill divided by the pages you actually ran. The same AWS instance that reads a page for $0.45 when saturated reads it for $11.75 at 50,000 pages a month and $58.75 at 10,000, because a GPU bills 730 hours a month whether documents arrive or not. Break-even against the $1.50 cloud rate sits around 391,700 pages a month on the cheapest adequate card. The full per-card arithmetic is on the self hosted OCR cost reference, and the published open-model figures are sourced and repriced on the olmOCR pricing page.
No single vendor is cheapest at everything, and the winner changes completely between reading text and extracting fields. Here is who actually wins each job at the July 2026 rates.
| The job | Cheapest option | Per 1,000 pages | Why |
|---|---|---|---|
| Read plain text off a scan | Azure Read / AWS Detect Text / Google OCR / Vision | $1.50 | A commodity. All three clouds match to the cent, so price is not a reason to choose. |
| Get tables and structure | Azure Layout or Google Layout Parser | $10.00 | AWS charges $15 for Tables, half again as much for the same job. |
| Read a standard business form | Azure or Google prebuilt | $10.00 | AWS has no invoice model as such, only Analyze Expense at $10. |
| Extract fields from a custom layout | Mistral Document AI | $5.00 | Six times cheaper than Azure or Google custom extraction at $30. |
| Route a mixed batch to the right model | Azure document classifier | $3.00 | Google charges $5, AWS offers no classifier at all. |
| Everything, at millions of pages | Azure Read commitment tier | $0.45 | The lowest verified rate from any major vendor. Needs 16M pages a month. |
Microsoft's marketing pricing page shows pay-as-you-go. It does not show the prepaid commitment tiers, which are a materially different price list. We pulled these from the official Azure Retail Prices API, the same feed the portal bills from, and re-verified every rate on July 23, 2026.
Two things stand out. The discount is steep on plain OCR, up to 70% off, and shallow on prebuilt models, 25% at a million pages a month. And the entry Read commitment starts at 500,000 pages a month, so most teams cannot reach it. If you process 250,000 pages a month you pay $375 pay-as-you-go, and the 500,000-page commitment also costs $375. At that volume the commitment doubles your allowance for free, and Microsoft never shows you.
The tier boundaries matter as much as the rates. AWS and Azure both halve basic OCR above one million pages a month, while Google Enterprise Document OCR does not discount until above five million, so the three vendors that list identically at $1.50 are furthest apart in exactly the band where most high volume pipelines sit. We modeled the whole curve from 100,000 to 10 million pages on high volume OCR API pricing.
Full Azure Document Intelligence pricing breakdown| Meter | Pages a month | Monthly fee | Per 1,000 pages | Off list |
|---|---|---|---|---|
| Read | 500,000 | $375 | $0.75 | 50% |
| Read | 2,000,000 | $1,200 | $0.60 | 60% |
| Read | 8,000,000 | $4,200 | $0.53 | 65% |
| Read | 16,000,000 | $7,200 | $0.45 | 70% |
| Prebuilt | 100,000 | $900 | $9.00 | 10% |
| Prebuilt | 1,000,000 | $7,500 | $7.50 | 25% |
| Custom extraction | 100,000 | $2,400 | $24.00 | 20% |
| Custom extraction | 1,000,000 | $18,000 | $18.00 | 40% |
Pages above the commitment bill at the same discounted rate, so the effective price per 1,000 pages holds whether you land under or over. Connected-container tiers run a further 10% to 20% below these.
Teams moving from an LLM provider assume batch processing is cheaper. On the cloud OCR services it is not. Azure's retail price feed meters batch analysis at exactly the same rate as a synchronous call: batch Read is $1.50 per 1,000 pages, batch Layout $10, batch prebuilt $10, batch custom extraction $30. AWS Textract and Google Document AI have no batch discount either. Batch on those platforms buys throughput and convenience, not a lower bill.
The two vendors that do discount asynchronous work are the LLM providers. Mistral cuts batch inference by 50%, taking Document AI from $5 to $2.50 per 1,000 pages. Gemini Batch is also 50% off. If you are processing an archive overnight rather than a document a user is waiting on, that is a real lever, and it only exists outside the big three clouds.
Each number above buys one thing: an API response. It does not buy a working document pipeline. If you are comparing a cloud rate against a product price, these are the lines that are missing from the cloud column.
A mixed batch has to be sorted before the right model reads it. Azure charges $3 per 1,000 pages for that, Google $5, and AWS does not offer it at all.
No API refunds a wrong read. Every uncertain field becomes correction time, and a cheap rate with a high correction rate can cost more in total than an expensive one that needs no cleanup.
Totals that reconcile, dates that parse, vendors that match your master data. That logic is yours to write and maintain against every layout change.
A JSON response is not an entry in your accounting or ERP system. The mapping and the sync are a project, not a line on a pricing page.
Google charges $0.05 an hour per deployed custom processor version, roughly $438 a year, whether or not you send it a single page. Azure charges nothing for a deployed model.
This is usually the biggest number on the page, and it never appears on a pricing page. Count it before you compare a rate against a product.
DocuOCR plans work out to roughly $14 to $20 per 1,000 pages, which is about ten times a cloud OCR rate and we are not going to pretend otherwise. The difference is what is in the box. A cloud API returns a response and leaves you the classification, the review screen, the validation and the export. DocuOCR ships those. If you have engineers, steady volume and time to build the pipeline, the cloud APIs are genuinely cheaper per page and you should use them. If you want the workflow already running, count the build before you compare the rates.
The buyer pillar: which OCR API to actually pick, not just what it costs.
Every meter, the free F0 tier, and the full commitment tier table.
Microsoft's newer service: $1 per 1,000 pages OCR and $0.01 on digital files.
Why Forms at $50 and Forms plus Tables plus Queries at $70 are the expensive lines.
The 10-page block rule and the idle hosting fee that accrues on an unused processor.
Every vendor at enterprise volume, with the discount tiers applied.
How the billing unit, not the rate, decides who is cheapest for you.
The cheapest published structured extraction rate, at $5 per 1,000 pages.
What the token rates actually work out to per 1,000 pages.
Every Claude model converted to dollars per 1,000 pages, and the resolution-tier trap.
GPT-5.4-nano at $1.67 per 1,000 pages, and the multiplier that halves its discount.
Every language model in one table, against the $1.50 dedicated OCR baseline.
Real AWS GPU rates and the volume each card needs to beat a paid API.
The $176 per million figure sourced, repriced on AWS, and checked at real volume.
The real curve from 100,000 to 10 million pages, and where each tier boundary sits.
Same $1.50 rate, three different answers to whether your document is stored.
What ZDR covers on Claude, Mistral and OpenAI, and the upload step it excludes.
Who actually ships containers, and why Azure on prem runs 2023 models.
Per document type, where the block minimums and per-doc rates bite hardest.
The cheapest API in the table is worthless if it reads your documents wrong, because every bad field costs more to fix than the page cost to process. Upload one of your real documents, look at the fields that come back, and pick on output rather than on a price list.
