How Much Does It Cost to OCR 1 Million Pages? (2026 Rates)

Reading 1 million pages of plain text costs $1,500 at list rates on any of the three big clouds, and as little as $600 once the volume tiers kick in. Extracting structured fields from the same million pages costs between $5,000 and $70,000. Here is the math, per vendor, at July 2026 rates.

Last updated July 12, 2026. Cloud rates re-verified that day against Microsoft's official Azure Retail Prices API, the AWS Textract pricing page and the Google Document AI pricing page.

Running plain OCR on 1 million pages costs $1,500 at list price on Azure, AWS or Google, and $600 to $750 a month once you clear the volume or commitment tiers. Extracting structured fields from those same million pages costs far more: $5,000 on Mistral Document AI, $10,000 on an Azure or Google prebuilt model, $20,000 to $30,000 on custom extraction, and $55,000 to $70,000 on AWS Textract Forms. The gap between reading text and understanding it is between 4x and 47x, and it is the single decision that sets your bill at this volume.

OCR cost for 1 million pages, by vendor and job

At a million pages a month you are past the first volume tier at every vendor, so list price is the wrong number to plan with. This table applies the discount tiers that actually apply at that volume.

Vendor and model	Rate at 1M pages	Cost for 1M pages	What you get
Azure Read, 2M commitment tier	$0.60 per 1,000	$1,200 (for 2M pages)	Text only
Azure Read, pay-as-you-go	$1.50 per 1,000	$1,500	Text only
AWS Detect Document Text	$1.50 per 1,000	$1,500	Text only
Google Enterprise Document OCR	$1.50 per 1,000	$1,500	Text only
Mistral Document AI	$5.00 per 1,000	$5,000 ($2,500 batched)	Structured output
Azure prebuilt, 1M commitment tier	$7.50 per 1,000	$7,500	Labeled fields, standard forms
Azure or Google prebuilt, list	$10.00 per 1,000	$10,000	Labeled fields, standard forms
Azure Layout / Google Layout Parser	$10.00 per 1,000	$10,000	Tables and structure
AWS Textract Tables	$15.00 per 1,000	$15,000	Tables
Azure custom, 1M commitment tier	$18.00 per 1,000	$18,000	Fields on your own layouts
Azure or Google custom extraction	$20.00 per 1,000 (above 1M)	$20,000 to $30,000	Fields on your own layouts
AWS Textract Forms	$40.00 per 1,000 (above 1M)	$40,000 to $50,000	Key-value pairs
AWS Forms + Tables + Queries	$55.00 per 1,000 (above 1M)	$55,000 to $70,000	Everything AWS offers

Read the top and bottom rows together. The same million pages costs $1,200 or $70,000 depending only on what you ask the API to do with them. Before you negotiate a rate, make sure you are buying the cheapest job that answers your question.

What is the cheapest way to OCR 1 million pages?

The cheapest way to read 1 million pages of plain text is a committed cloud tier. Azure's Read commitment at 2 million pages a month is $1,200, which works out to $0.60 per 1,000 pages, and the 16-million-page tier reaches $0.45 per 1,000, the lowest verified rate any major vendor publishes. AWS and Google both step plain OCR down to $0.60 per 1,000 pages at high volume too, so all three converge. At a million pages, plain OCR is a commodity priced within pennies of itself across the industry.

The Azure commitment tiers most teams never see

Microsoft's pricing page shows pay-as-you-go rates. It does not show the prepaid commitment tiers, which are a separate and much cheaper price list. They are exposed through the Azure Retail Prices API, the feed the portal bills from. These are the rates as of July 12, 2026:

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	16,000,000	$7,200	$0.45	70%
Prebuilt	1,000,000	$7,500	$7.50	25%
Custom extraction	1,000,000	$18,000	$18.00	40%

Pages above the commitment bill at the same discounted rate, so you are not punished for overshooting. The catch is the floor: the smallest Read commitment is 500,000 pages a month, which is more than most teams process, so below that the tiers simply are not available to you.

Does batch processing make 1 million pages cheaper?

Not on the big three clouds. Azure meters batch analysis at exactly the same rate as a synchronous call, and neither AWS Textract nor Google Document AI offers a batch discount. Batch on those platforms buys throughput, not a lower bill. The vendors that do discount asynchronous work are the LLM providers: Mistral cuts batch inference by 50%, taking Document AI on a million pages from $5,000 to $2,500, and Gemini Batch is also half price. If you are digitizing an archive overnight rather than answering a user in real time, that is the one lever that genuinely moves the number.

What the API bill leaves out at this volume

A million pages a month is a production system, not a script, and the API line is not the whole cost. Count these before you take a number to a budget meeting.

Classification. A million mixed pages have to be sorted before the right model reads them. Azure charges $3 per 1,000 pages for a classifier, Google $5, and AWS has none, so on AWS you build it.
Review. At 99% field accuracy, a million pages still produces thousands of wrong fields. Someone fixes them. That headcount usually costs more than the API.
Reprocessing. Failed or low-confidence pages get retried, and Azure and AWS bill each attempt. Only Google skips billing on 4xx and 5xx errors.
Storage and egress. A million scanned pages is real object storage plus the traffic to move it.
Idle hosting. Google charges $0.05 an hour for each deployed custom processor version, about $438 a year, whether it processes a page or not.

Teams that push extracted invoice data straight into finance usually pair the extraction step with accounts payable automation software rather than dropping a million JSON files into a bucket and calling it done. The extraction is the cheap half. The workflow around it is where the volume actually hurts.

How much does it cost to OCR 1 million pages once, not monthly?

A one-time backfill of 1 million pages is the same $1,500 of plain OCR at list price, but you almost certainly cannot use a commitment tier for it, because those are monthly subscriptions sized to steady volume. Signing a 2-million-page Azure Read commitment at $1,200 to process a single million-page archive means paying $1,200 for a month you will not repeat, which is still cheaper than $1,500 pay-as-you-go, but only just. For a genuine one-off, take the $0.60 per 1,000 volume tier that Azure, AWS and Google all apply automatically above their thresholds, or run it through Mistral batch at $2 per 1,000 pages if you want structure rather than text.

Is running your own OCR model cheaper at a million pages?

Rarely, and later than people expect. Open-weight models like DeepSeek-OCR are free to license and cost GPU hours to run. A 24/7 L40S on a community cloud runs roughly $722 a month, which only undercuts Azure Read at around 480,000 pages a month, so at a million pages self-hosting does start to win on raw compute. What it does not include is the engineer keeping the inference service alive, the accuracy tuning, and the fact that an open OCR model gives you text, not validated fields. Self-hosting is a real option at eight-figure page volumes and a distraction below that.

The short version

Plain OCR on a million pages is $600 to $1,500 and is not worth optimizing. Structured extraction on a million pages is $5,000 to $70,000 and is worth an afternoon of arithmetic. Pick the cheapest model that answers your question, check whether your volume clears a commitment tier, and then spend your remaining energy on accuracy, because at a million pages the correction time costs more than the API ever will.

For the full normalized rate table across every vendor, see OCR pricing per 1,000 pages. For a single vendor, read Azure Document Intelligence pricing, AWS Textract pricing or Google Document AI pricing. If you are choosing rather than costing, the OCR API pricing comparison pillar covers the decision, and enterprise document OCR covers what changes at this scale.