Document AI cost per page in 2026: 34 prices compared

What does document AI cost per page in October 2026?

Plain OCR costs $1.50 per 1,000 pages on Amazon Textract, Azure and Google at first-tier list prices, and $4 on Mistral OCR 4.1. Field extraction runs from about $0.35 per 1,000 one-page invoices on Gemini 2.5 Flash-Lite to $100 on Google's Invoice parser.

We read the pricing pages and pricing data of seven providers on October 30, 2026 and recorded 34 price points: 24 services priced per page or per document, and 10 AI models priced per token, which we converted to a cost per 1,000 pages with the arithmetic shown. The result is the table we use to pick an extraction service when we scope an AI document processing build. Every figure links to the provider's own page.

Three findings stand out:

• Reading is cheap, labelling is not. All three cloud OCR services charge the same $1.50 per 1,000 pages. Named fields cost from $4 (Textract Layout) to $100 (Google's Invoice parser) on the same providers.
• Small AI models undercut prebuilt parsers. In our arithmetic five AI models extract a one-page invoice for under $2 per 1,000, below the $10 prebuilt invoice models on Textract and Azure.
• The model tier matters more than the vendor. Inside one vendor the spread is wide: Claude Haiku 5.5 works out at about $1.17 per 1,000 invoices and Claude Sonnet 5.5 at about $23.35, on the same assumptions.

How did we build this dataset?

We read each provider's own pricing page or pricing feed on October 30, 2026, recorded the first-tier US list price per 1,000 pages, and converted token prices with one stated set of assumptions. We did not use vendor marketing figures, surveys or third-party price lists.

Rules we applied. First pricing tier only, unless a row says otherwise. No free tiers in the main numbers. Per-document prices (Google's Invoice and Expense parsers) are shown per 1,000 one-page documents. AI model rows use the same invoice job as our service page, so the two pages agree.

Assumptions for the AI model rows, labelled because the providers do not publish one fixed number per page:

• Each page is sent with an 800-token instruction and field list, and the model returns about 600 tokens of structured fields. No extended thinking.
• Gemini 2.5: 258 tokens per page, from Google's statement that "each document page is equivalent to 258 tokens".
• Gemini 3.x: 560 tokens per page, the default PDF value in Google's media resolution table for Gemini 3 models. Google says Gemini 3 does not charge for native text extracted from PDFs.
• Claude: Anthropic says each page uses 1,500 to 3,000 text tokens plus image tokens; we used 1,500 text tokens. For the image we used each model's maximum visual tokens, 1,568 for Haiku 4.5 and 4,784 for Claude 4.7 and later models, so these rows are an upper bound. For the newer models we also added the roughly 30 percent more tokens Anthropic says its newer tokenizer produces for the same text.
• OpenAI: 1,500 text tokens plus a 3,000-token page image, which is the 2,500-patch high-detail budget times the 1.2 multiplier OpenAI lists for its current vision models. The multiplier table does not name the GPT-6 Luna, Terra or Sol models, so this is our assumption for them.

What does plain OCR cost per 1,000 pages?

$1.50 per 1,000 pages on Amazon Textract, Azure Read and Google Enterprise Document OCR, falling to $0.60 at high volume. Mistral OCR 4.1 lists $4 per 1,000 pages, and Textract's Layout feature, which adds structure, lists $4.

Sources: Amazon Textract pricing, Azure Retail Prices API, Google Document AI pricing, Mistral OCR 4.1, all read October 30, 2026. Textract Layout and Mistral OCR return document structure as well as text, so the $4 rows buy more than plain OCR; this table compares price only.

What do structured extraction services cost per 1,000 pages?

Structured extraction costs $10 per 1,000 pages for prebuilt invoice and expense models on Textract and Azure, $15 to $70 for Textract's tables, queries and forms features, $30 for custom extractors on Azure and Google, and $100 per 1,000 one-page documents on Google's Invoice parser.

Two pricing details change real bills. Google notes that "the cost to call the custom extractor is the same whether you use a generative AI-powered processor version or custom model", so moving to a newer Gemini-based version does not change the per-page price. And Google's per-document parsers charge the same for a one-page and a ten-page invoice, which favours long documents and penalises short ones.

What do AI vision models cost per 1,000 invoice pages?

On our assumptions, from about $0.35 per 1,000 one-page invoices on Gemini 2.5 Flash-Lite to about $23.35 on Claude Sonnet 5.5. Five models land under $2, which is below the $10 prebuilt invoice models on Textract and Azure. Batch processing halves every figure.

Sources: Gemini API pricing, Gemini document processing, Gemini media resolution, Claude pricing, Claude PDF support, Claude vision, OpenAI pricing, OpenAI images and vision, read October 30, 2026. Batch prices are each provider's listed batch rates, which are half the standard rates for every model in the table.

Two dated changes to watch. Google lists Gemini 3.6 Flash at $0.75 input and $3.75 output through December 31, 2026, and at $1.50 and $7.50 from January 1, 2027, which doubles that row to about $6.54. And Claude Haiku 5.5's $0.10 input rate applies to prompts up to 100,000 tokens; a single long document sent in one request is priced at $0.50 input.

How do volume tiers and commitments change the bill?

At 5 million OCR pages a month, Textract and Azure Read cost about $3,900 each, Google about $7,500, and Mistral OCR 4.1 about $20,000 at list price. Batch APIs halve AI model costs, and Azure and Google sell commitment plans that cut per-page rates.

Google's price only drops to $0.60 above 5 million pages, while Textract and Azure drop at 1 million, which is why Google costs nearly twice as much at this volume despite the same starting price.

Commitment pricing, read from the same sources:

• Azure commitment tiers. The Retail Prices API lists, for example, a Read commitment of 2 million pages a month at $1,200 (an effective $0.60 per 1,000 pages) and a prebuilt commitment of 100,000 pages at $900 a month (an effective $9 per 1,000), with overage rates for pages above the commitment.
• Google savings plans. Google's pricing page shows Gemini Enterprise Flexible Savings Plan prices: the custom extractor drops from $30 to $27 on a 1-year plan and $24 on a 3-year plan, and the Invoice parser from $0.10 to $0.09 and $0.08 per document.
• Batch APIs. Gemini, Claude and OpenAI batch rates are half their standard rates. Mistral's pricing page says batch processing reduces prices by 50 percent in general but does not list a separate OCR batch price, so we left OCR at list.

What does this dataset prove, and what does it not?

It proves what each provider charged at list price on October 30, 2026, and what those prices mean for one defined invoice job. It does not measure accuracy, latency or data terms, and token-based rows depend on assumptions that can be off for dense pages.

What it shows:

• The published price of 34 document AI options on one date, from primary sources.
• That the three big cloud OCR services start at the same price, and that their volume breakpoints differ.
• That on our assumptions several small AI models cost less per invoice than prebuilt invoice parsers.

What it does not show:

• Accuracy. A cheaper model that gets more fields wrong costs more in review time. We test on your documents before choosing.
• Exact token counts. Dense pages, multi-page tables and photos can use far more tokens than our assumptions. The Claude rows use the maximum image size and are an upper bound; real PDF pages are often smaller.
• Features. Prebuilt parsers return named fields and confidence scores without prompt design. AI models need a schema, and not all of them return per-field confidence.
• Data terms, regions and taxes. Prices vary by region and currency, and data handling differs by provider.
• Everything around extraction. The automation platform, storage, the accounting system and the build are separate costs.

How should you use these numbers when choosing a service?

Start from your documents, not the cheapest row. Pick the cheapest options that support your languages and document types, test them on 20 to 50 real documents, and compare the total of extraction cost plus review time. Re-check prices before you commit, because they change.

A short way to read the tables:

• Searchable text only: the $1.50 OCR services. At high volume, the 1 million page breakpoint on Textract and Azure matters.
• Standard invoices and receipts with little setup: the $10 prebuilt models on Textract or Azure.
• Varied layouts, custom fields or high volume: a small AI model with a fixed schema, ideally through a batch API when same-day results are fine.
• Handwritten forms: language support decides first; our FAQ on whether AI can read handwritten documents compares it by provider.

The price of extraction is usually the smaller part of the decision. Our OCR vs AI extraction FAQ explains which layer each document needs, and our guides to QuickBooks AP automation and Xero invoice processing show where the extracted data goes. The platform side is covered in our n8n and Zapier pricing research.

More questions about document AI pricing

What is the cheapest OCR API per page?

At first-tier list prices on October 30, 2026, Amazon Textract Detect Document Text, Azure Read and Google Enterprise Document OCR all cost $1.50 per 1,000 pages. Above 1 million pages a month, Textract and Azure drop to $0.60; Google drops to $0.60 only above 5 million. Google also gives the first 1,000 pages a month at $0.

How much does Mistral OCR cost per page?

Mistral's OCR 4.1 model page lists $4 per 1,000 pages, which is $0.004 per page, and $5 per 1,000 annotated pages, the option that returns structured fields. Mistral's changelog says OCR 4.0 retired on September 30, 2026, with users moved to OCR 4.1 "at the same price". Mistral does not list a separate OCR batch price.

How much does Amazon Textract cost per page?

In US East (N. Virginia) on October 30, 2026: $0.0015 per page for Detect Document Text, $0.01 for Analyze Expense, $0.015 each for Tables or Queries, $0.05 for Forms and $0.07 for Forms, Tables and Queries together, for the first million pages a month. Prices drop above one million pages. Layout is $0.004 and is included free with Forms, Tables or Queries.

Is it cheaper to use ChatGPT or Claude than a document AI service?

For short, sparse pages, often yes. On our assumptions GPT-6 Luna costs about $0.83 and Claude Haiku 5.5 about $1.17 per 1,000 one-page invoices, below the $10 prebuilt invoice models. Larger models cost more: GPT-6.1 Sol about $16.60 and Claude Sonnet 5.5 about $23.35. Dense pages raise the token count, so measure on your own documents.

How many tokens is one PDF page?

It depends on the provider. Google says a page is 258 tokens on Gemini 2.5, and its Gemini 3 media resolution table lists 560 for PDFs at the default setting. Anthropic says 1,500 to 3,000 text tokens per page plus image tokens. OpenAI sends extracted text plus a page image, which at high detail can reach about 3,000 image tokens.

Do free tiers make a difference?

For testing, yes; for production, rarely. Google gives the first 1,000 OCR pages at $0, Textract gives new AWS customers 1,000 Detect Document Text pages and 100 Analyze Expense pages a month for three months, and Azure has a free tier. At a few thousand pages a month the paid price is already the larger part of the bill.

How often do these prices change?

Often enough to re-check before every build. Within the period we looked at, Google listed a doubling of Gemini 3.6 Flash prices from January 1, 2027, and Mistral retired OCR 4.0 on September 30, 2026. We date every price on this page and update the table when a provider changes one.

Sources and further reading

• AWS: Amazon Textract pricing (US East, N. Virginia), checked October 30, 2026
• Microsoft: Azure Retail Prices API, Azure Document Intelligence, East US, checked October 30, 2026
• Google Cloud: Document AI pricing, checked October 30, 2026
• Mistral: OCR 4.1 model page, changelog and pricing FAQ, checked October 30, 2026
• Google: Gemini API pricing, document processing and media resolution, checked October 30, 2026
• Anthropic: Claude pricing, PDF support and vision, checked October 30, 2026
• OpenAI: API pricing, images and vision and file inputs, checked October 30, 2026