OCR or invoice extraction? Price the document result you need
Plain text extraction and a structured invoice result are two different jobs, and they come with very different price tags. AWS's selected Oregon Detect Document Text example uses $0.0015 per page, so 10,000 pages come to $15 outside the free tier. Google's Invoice Parser works differently: it bills $0.10 for up to ten pages in each document. Let's sort out which one you actually need before you start budgeting.
This guide starts with the document fields you need. For a high-volume parser comparison that includes Mistral, use the million-page OCR guide. Mistral OCR vs AWS Textract vs Google Document AI: Real Document AI Cost per 1 Million Pages in 2026.
Define the output before comparing a thousand pages
The easiest way to blow a document-processing budget is to treat every document task as basic OCR. If all you need is raw text strings and their bounding box coordinates, basic OCR does the job cheaply. But if you want structured invoice extraction (supplier names, invoice dates, tax rates, line items, and totals mapped to standard schema keys), you're in specialized document model territory, and those carry very different rate cards. So start with the business output you want, then pick the processor that delivers it.
| Selected task | Example billing unit | Budget implication |
|---|---|---|
| AWS Detect Document Text, Oregon example | $0.0015 per page in the first million, outside free tier | 10,000 pages cost $15 |
| AWS Analyze Document forms + tables, same selected tier | $0.05 forms + $0.015 tables per page | 1,000 pages cost $65; OCR is included |
| Google Invoice Parser | $0.10 for each up-to-ten-page block within a document | 100 separate one-page invoices cost $10 |
| Azure Document Intelligence | Selected Read, layout or prebuilt extraction SKU and region | Quote the actual processor, not a generic OCR rate |
Document boundaries can change the calculation
Here's a quirk worth knowing: the vendors don't all count the same thing. Google Document AI's specialized Invoice Parser charges $0.10 per document block (up to 10 pages per block), so an 11-page invoice counts as two blocks ($0.20). That means processing 100 separate one-page invoices costs $10.00 (100 blocks), while processing a single 100-page batch document consumes 10 blocks ($1.00). The practical catch is that page costs don't scale linearly, so it's worth checking file boundaries before you trust a per-page estimate.
Flip it around and the math moves the other way. If you process 100 eleven-page invoices on Google Document AI, each document consumes two blocks, bringing the bill to $20.00. My advice: preserve natural document boundaries in your ingestion pipeline, and double-check your provider's supported maximum page limits while you're at it.
Avoid paying for the same OCR twice in the estimate
This one trips up a lot of spreadsheets. AWS documentation makes a helpful distinction: Analyze Document calls automatically include raw text OCR, whether you request forms, tables, or queries. Analyze Expense and Analyze ID also include OCR within their per-page fee. So if you run Analyze Document with forms and tables ($0.05 + $0.015 = $0.065/page), don't add a separate $0.0015 Detect Document Text line item. You'd be paying for OCR twice, but only on paper.
Over on Microsoft Azure, Document Intelligence separates its high-speed Read model ($0.0015/page) from Prebuilt Invoices and Receipts ($0.01/page) and custom extraction. What to keep in mind: if you quote Azure Read when your app really needs structured invoice parsing, your budget will be off by an order of magnitude. Google follows a similar pattern, separating Enterprise OCR from its Form and Invoice processors.
Test the documents that cause manual work
Please don't base a document budget on a sample of clean digital PDFs. They're lovely, and they're not what shows up in production. Build a real-world test batch with blurry mobile camera photos, skewed scans, wrinkled receipts, multi-page tables, and receipts with non-standard currencies or date formats. Then compare the parsed JSON values against manual ground truth, paying special attention to decimal separators, currency symbols, and invoice totals.
Track your team's human review time during the pilot, too. If human operators must manually review and correct 10% of invoices, and each correction takes five minutes, 10,000 invoices demand roughly 83 hours of clerical labor. At $25/hour, that labor costs over $2,000, which dwarfs the cloud OCR API bill. Accuracy that eliminates manual review is worth far more than a tiny per-page API discount.
Keep retries, storage and validation in the workflow
Plan a clear fallback path for documents that fail extraction or come back with low confidence scores. Remember to factor in S3 or Google Cloud Storage fees, egress bandwidth, database ingestion, and any downstream LLM structuring. And store the original PDF files alongside the extracted JSON, so human auditors can verify disputed figures later (future you will be grateful).
Our worksheet below starts with AWS Oregon Detect Document Text at $0.0015/page ($15.00 for 10,000 pages). If your workflow requires forms and tables, update the rate to $0.065; if you're using block-priced parsers, calculate the billable blocks before entering your volume.
Recalculate a flat per-page quote
Starts with the AWS Oregon Detect Document Text example outside the free tier and within its first-million-page band. This does not automatically apply volume tiers or document-block billing. Enter only applicable free pages and billable retries.
Budget the complete document-processing job →