AI Document Extraction Pricing: Per Page vs Per Token Cost Math
You have three quotes and no clean way to compare them. One is priced per million tokens. One is priced per 1,000 pages. One is a monthly credit bundle you are supposed to estimate against before processing anything. The unit prices differ by three orders of magnitude, which is the first clue that the units are not measuring the same thing.
Short Answer
Per-token pricing is the cheapest per successful page and the most expensive per failed one, because every retry, reasoning pass, and schema-change rerun bills again at full rate. Per-page pricing is predictable on volume but usually says nothing about whether the extracted values are right. Credit bundles combine the volume risk of one with the opacity of the other. The only model that moves rework risk off your side of the table is per-page pricing with an accuracy commitment attached, because a wrong value then costs the vendor money instead of costing you a week.
| Per token | Per page (API) | Credits / subscription | Per page + accuracy commitment | |
|---|---|---|---|---|
| You are billed for | Input, output, and reasoning tokens | Pages submitted | Credits drawn per page by mode | Pages delivered |
| Who pays for a retry | You | You | You | Vendor |
| Who pays for a wrong value | You, twice: rerun plus review | You | You | Vendor |
| Blank or failed page | Billed | Billed | Billed | Not billed |
| Cost knowable before you start | No | Yes, if you know the page count | Only inside the bundle | Yes |
| Needs engineering time | Yes | Yes | Usually | No |
| Best fit | High volume, error-tolerant output | Uniform, native-text documents | Steady predictable mix | Mixed quality, correctness-critical |
What the Published Rates Actually Are
Real numbers as of August 2026, so you can see the spread before deciding which spread matters.
Per page. Google Document AI charges $1.50 per 1,000 pages for Enterprise Document OCR, $10 for Layout Parser, and $30 per 1,000 pages for Form Parser and Custom Extractor. AWS Textract charges $1.50 per 1,000 pages for plain text detection, $15 for Tables, $15 for Queries, $10 for Analyze Expense, and $50 per 1,000 pages for Forms. Azure AI Document Intelligence bills per 1,000 pages analyzed across its Read, Layout, prebuilt, and custom models.
Per credit. LlamaParse sells credits at $1.25 per 1,000, which is $0.00125 each. Its tiers consume 1 credit per page for Fast, 3 for Cost-effective, 10 for Agentic, and 45 for Agentic Plus. Layout extraction adds 3 credits per page. That is a 45x range on the same physical page depending on which mode the document turns out to need.
Per token. Gemini 3.5 Flash-Lite is $0.30 per million input tokens and $2.50 per million output. Gemini 3.5 Flash is $1.50 and $9.00. A page of structured JSON output runs near 900 tokens, so one clean pass on a cheap model costs a fraction of a cent.
The gap between the cheapest and most expensive published unit is roughly 40x, and not one of these numbers tells you what share of the extracted values will be correct.
Have a messy PDF? Upload 1-3 sample pages and we will tell you if it is clean, OCR-heavy, or needs human QA.
Per Token: The Meter Runs Fastest on Failure
Token billing has an inverted cost curve. A page that parses correctly on the first attempt is close to free. A hard page gets retried after failing schema validation, escalated to a larger model, re-read with a longer prompt, and then handed to a person anyway. You pay for every attempt.
Three mechanisms drive that:
- Output tokens dominate. On Flash-class models the output rate is six to eight times the input rate, and a full-page JSON table is mostly output. Shortening what the model writes saves more than shortening what it reads.
- Reasoning tokens bill at output rates. A model hedging on a blurry figure thinks longer, and the thinking is billed like the answer.
- A schema change reruns the corpus. Add one field in month two and every page already processed gets read again at full price.
Claude's PDF handling illustrates the input side: each page is converted to an image and its text layer extracted, with the documentation putting the text alone at 1,500 to 3,000 tokens per page before image tokens are added. Richer signal on stamped forms, higher input cost per page.
Per Page: Predictable Until You Read the Page Definition
Per-page pricing solves forecasting, not correctness. Before you trust a quote, find out what the vendor counts as a page. Google counts each PDF page, each image in a TIFF, each tab in a spreadsheet, and every 3,000 characters of a Word or HTML file as one page. Textract bills per feature, so running Tables and Forms on the same page is $15 plus $50 per 1,000, not $50.
The more important detail: a page that returns unusable output bills exactly the same as a page that returns a perfect table. Revenue is indexed to volume submitted, not value delivered. Fair when your documents are uniform and machine-generated. Bad when a quarter of your corpus is a photocopy of a fax.
Credits: Predictability That Evaporates on Contact
Credit systems are per-page pricing with the page price hidden behind a mode selector. Which tier a document needs is precisely what you cannot know at quote time. Buy 40,000 credits for $50, discover that half your scans need the agentic tier rather than the fast one, and your effective per-page cost moves by more than an order of magnitude while the monthly allowance disappears in week one.
Run One Real Job Through All Three
Assume 5,000 pages, a mix of native PDFs and scans, tables on most pages, and a 15% rate of pages that need a value corrected.
| Approach | Unit rate | First-pass cost | Cost of reprocessing 15% |
|---|---|---|---|
| DIY on Gemini 3.5 Flash-Lite | ~$0.0023/page | ~$12 | ~$2 in tokens, plus engineering time |
| LlamaParse Agentic | $0.0125/page | $63 | ~$9 |
| Google Custom Extractor | $30/1,000 | $150 | ~$23 |
| Textract Tables plus Forms | $65/1,000 | $325 | ~$49 |
Look at the right-hand column. The entire spread of reprocessing charges is under $50. Now price the part nobody quotes: 750 pages that need a human to open the source PDF, find the wrong value, and fix it. At four minutes each that is 50 hours. At any realistic loaded rate, that line exceeds every API charge in the table combined, several times over.
Arguing about per-page versus per-token unit rates is close to a rounding error. The question that decides project cost is who is responsible for the 15%.
Why an Accuracy Commitment Changes the Incentive
Under both per-token and per-page billing, the provider is paid the same whether the numbers are right. Attach a correction commitment to a per-page price and the incentive inverts. A vendor who has to redo bad pages for free will do things that cost them margin up front: cross-foot totals against line items, check that an invoice's rows sum to its stated total, flag low-confidence fields instead of letting a plausible guess through, and put a person on the pages the model was least sure about. They will also decline jobs they cannot do well, which is the most useful signal you will get in a sales conversation.
Push on what the commitment says. Field-level accuracy on a named list of fields, measured against a sample you select after delivery, with rework at the vendor's cost, is a commitment. "99% OCR accuracy" measured per character is not, because one wrong digit in an amount column still scores above 99.9% and still breaks the spreadsheet.
When Per Token Is Genuinely the Right Choice
If you have engineers, volume in the tens of thousands of pages a month, documents from a consistent source, and a downstream use that tolerates noise, build the pipeline and pay per token. Search indexes, retrieval corpora, classification, and triage queues absorb occasional errors without anyone getting hurt, and at that volume the token bill is a small fraction of any managed alternative. Same for a one-time job small enough to check yourself.
If your documents are uniform, native-text, and well-formed, a per-page API from one of the cloud providers is cheap and adequate. Take it and spend your money elsewhere.
Managed per-page pricing with an accuracy commitment earns its premium in a narrower band: mixed-quality sources, scanned and stamped and rotated pages, values that land in accounting records or filings, and no engineering time to spend on a pipeline.
What to Ask Before You Sign
- What counts as a page: spreadsheet tabs, multi-image TIFFs, character counts in Word files.
- Are failed, blank, or unreadable pages billed.
- Are retries and second passes billed to me or absorbed.
- What is the accuracy definition, at what level, on which fields.
- What is the remedy when accuracy is missed: rework, refund, or an apology.
- Do unused credits roll over, and when do they expire.
- What happens to the price when I add a field after delivery.
Frequently Asked Questions
Is per-token pricing always cheaper than per-page pricing?
Per successful page, usually yes, often by 10x. Per project, frequently no. Token pricing excludes the engineering time to build the pipeline, the retries and reruns, and the human review of whatever the model got wrong. Comparing unit rates alone means comparing the smallest line item in the budget.
What is a fair per-page price for AI document extraction?
It depends on what is included. A raw OCR call is $0.0015 per page. A structured extraction API is $0.03 to $0.065. A managed service including human review, validation against document totals, and free correction sits well above that, because four minutes of a person's attention costs more than any of those API calls. Price the deliverable, not the unit.
How do I compare a token quote to a page quote?
Convert both to cost per correct field. Take 50 representative pages, run them through each option, and count fields that are right, wrong, and missing. Divide total cost including your review time by the number of fields you would have shipped untouched. That number is comparable across pricing models. Nothing on the vendors' pricing pages is.
Do accuracy commitments actually get paid out?
The payout matters less than the behavior change. A vendor carrying rework liability builds validation into the process and staffs review on the pages that need it, because it is cheaper for them than redoing the job. Read the remedy clause anyway: a commitment with no defined consequence is marketing copy.
Sources: Google Document AI pricing, AWS Textract pricing, Azure AI Document Intelligence pricing, LlamaParse pricing, Gemini API pricing, Claude PDF support
Ready to Convert Your Documents?
Stop wasting time on manual PDF to Excel conversions. Get a free quote and learn how DataConvertPro can handle your document processing needs with AI-assisted extraction and human verification.