Best Document Extraction APIs (PDF to JSON)

"PDF to JSON API" covers three genuinely different products, and comparing them on price alone produces nonsense answers. One group recognises characters. One group turns a document into structured text for a language model. One group returns named business fields. They differ in price by a factor of a hundred, and for good reason.

Here are nine of them, sorted by what they give back, with pricing normalised per 1,000 pages where the vendor publishes enough to do so. Checked August 2026.

Short version

Cheapest text at scale: AWS Textract, Google Document AI, or Azure Document Intelligence, all around $1.50 per 1,000 pages.

Best for LLM pipelines: Reducto, LlamaParse, or Unstructured, depending on whether you value table fidelity or free volume.

Bank statements specifically: the CapyParse API, which returns normalized transactions instead of text you have to interpret.

Three categories, not one

OCR primitives. AWS Textract, Google Document AI, Azure Document Intelligence. Decades of engineering behind them, priced like utilities, and they return words with coordinates plus optional table and form structure. Everything downstream is yours.

General parsers. Reducto, LlamaParse, Unstructured. Newer, built for retrieval and language model pipelines. They preserve document structure and hand you markdown or chunked JSON. Excellent for feeding a model, still not a ledger.

Domain specialists. Mindee, Nanonets, CapyParse. They return fields with meaning: invoice total, transaction amount, bill of lading number. You pay more per page and skip the part where you write the logic that decides which of the three numbers on a line is the amount.

Pick the category first. Comparing a $1.50 OCR call against a specialist API is comparing flour to bread.

Pricing, normalised

API	Category	Per 1,000 pages	Free tier
AWS Textract	OCR primitive	$1.50 text, $15 tables, $50 forms	3 months, limited
Google Document AI	OCR primitive	$1.50 OCR, $10 layout, $30 form parser	Varies by processor
Azure Document Intelligence	OCR primitive	$1.50 read, $10 layout, $30 custom	500 pages/mo, 2-page cap
LlamaParse	General parser	~$1.25 basic mode	10,000 credits/mo
Unstructured	General parser	$30	15,000 pages/mo
Reducto	General parser	$0.015 per credit, mode-dependent	15,000 credits
Mindee	Specialist	~$44	14-day trial
Nanonets	Specialist	$20 to $300 per block run	$50 credit
CapyParse	Specialist	$132 to $193	10 pages

Read that table sideways rather than top to bottom. The $1.50 rows return words. The bottom rows return a reconciled transaction list. If you are converting bank statements, the price difference buys back several weeks of building the layer in between, and if you are indexing research papers it buys you nothing at all.

OCR primitives

1. AWS Textract

The most feature-granular of the three, which is both the appeal and the trap. Text detection is $1.50 per 1,000 pages, tables $15, forms $50, and queries $15. Analyze Lending, aimed at loan packages including bank statements, is $70.

The trap is that these stack. Ask for text, tables, and queries on the same document and you pay for all three, so a careless implementation lands twenty times above the rate you quoted your finance team. The free tier lasts three months rather than being permanent.

Best for: teams already on AWS with high volume and a clear idea of which features they need.

2. Google Document AI

Enterprise Document OCR at $1.50 per 1,000 pages, Layout Parser at $10, and Form Parser or Custom Extractor at $30. Volume tiers cut those roughly by a third past a million pages.

The line item people forget is $0.05 an hour for each deployed custom processor version, charged whether or not you send it any documents. That is about $36 a month per processor sitting idle, which matters if you spin up one per document type.

Best for: Google Cloud shops with steady throughput.

3. Azure AI Document Intelligence

Read is $1.50 per 1,000 pages, Layout $10, prebuilt models $10, and custom extraction $30, with add-ons at $6. Layout returns table structure as well as text, which for financial documents is worth the step up from Read.

Know the free tier's real shape before you plan around it: F0 covers 500 pages a month but processes only the first two pages of any request. A ten-page statement returns two pages of results and no error explaining why.

Best for: Azure environments, and anyone who wants table structure without paying custom-model rates.

General parsers

4. Reducto

Reducto positions itself around table fidelity at the cell level, which is exactly the failure mode that ruins financial document pipelines. Parse, Extract, Edit, and Split APIs cover 30-plus file types.

Pay as you go starts with 15,000 free credits then $0.015 per credit, with a 20% discount for async jobs that can wait up to 12 hours. Growth and Enterprise add VPC and on-premise deployment. Credits per page depend on the parse mode, so run your own documents before extrapolating a monthly bill.

Best for: pipelines where a misread table cell has direct financial consequences.

5. LlamaParse

Part of LlamaCloud and closely tied to the LlamaIndex ecosystem, which makes it the path of least resistance if your retrieval stack is already built there. Free usage is 10,000 credits a month, Starter is $50 for 40,000, and Pro is $500 for 400,000. Credits convert at 1,000 for $1.25.

Basic parsing starts at one credit per page, so simple documents are close to $1.25 per 1,000 pages. The agentic modes that handle complex layouts cost substantially more per page, and that is where budgets drift.

Best for: RAG pipelines already using LlamaIndex.

6. Unstructured

The most generous ongoing free tier here: 15,000 pages a month, no card, resetting every month. After that it is $0.03 a page, and the bill caps at $3,000 a month with further pages free up to a million.

It handles a very wide range of file types and produces chunked, structured output aimed at embedding pipelines. There is also an open source library if you would rather run it yourself.

Best for: mixed-format corpora headed into a vector database.

Domain specialists

7. Mindee

Prebuilt models for invoices, receipts, bank statements, IDs, bills of lading, resumes, and contracts, with documentation good enough to have something working the same afternoon. Credits are counted per physical page at roughly $0.044, Starter is $44 a month and Pro is $116 on annual billing.

Enterprise requires a commitment of 500,000 credits a year, which is worth knowing before you build a roadmap around enterprise-only features.

Best for: teams wanting named fields for common document types with predictable per-page cost.

8. Nanonets

A workflow platform rather than a single endpoint. You compose blocks: classification, extraction, validation, custom Python. Simple blocks are $0.02 a run, standard AI $0.10, complex AI $0.30, with $50 of free credit and a published $100 a month entry plan.

Powerful when documents need routing before extraction. Harder to forecast than a flat per-page rate, because the cost of a document depends on how many blocks it traverses.

Best for: mixed document streams needing classification first.

9. CapyParse: bank statements, already reconciled

A narrow API on purpose. POST a PDF to /api/v1/conversions as multipart, a URL, or base64, poll for status, and fetch the result as JSON, JSONL, CSV, or markdown. Accounts inside a multi-account statement come back separated rather than merged.

Authentication is a team-scoped key sent as a bearer token, billing is one credit per page, and there is an MCP server at /mcp plus a stdio package if you want an agent doing the conversion. Results can be returned with provenance so a downstream agent can cite where a figure came from rather than asserting it.

Pros

Transactions, not text you must interpret
JSON, JSONL, CSV, and markdown output
MCP server for agent workflows

Cons

Financial documents, not general parsing
Higher per page than raw OCR
Asynchronous, so you poll for results

Best for: anything that ends with transactions in a database. See the developer docs.

Try the API on a real statement

10 free pages on signup, no card. Same credits work through the dashboard, the REST API, and MCP.

Read the API Docs

Four things to check before you commit

Model your real page mix

Take 100 representative documents, count the pages, and multiply. Headline rates assume single-feature calls on single-page files, and almost nobody has those.

Find the idle costs

Deployed processors, monthly minimums, and annual credit commitments all bill whether you send documents or not.

Ask what happens when it is wrong

Does the response tell you which fields were uncertain, or does everything come back with equal confidence? That determines whether you can automate the happy path.

Check the retention policy

If you are processing financial documents for other people, how long the vendor keeps files is a contractual question, not a technical one.

Frequently asked questions

What is the cheapest PDF to JSON API?

For plain text recognition, AWS Textract, Google Document AI, and Azure AI Document Intelligence all sit at about $1.50 per 1,000 pages. LlamaParse basic mode works out near $1.25 per 1,000 pages. Structured extraction costs considerably more: Google Form Parser and Azure custom extraction are both $30 per 1,000 pages.

Which document extraction API has the best free tier?

Unstructured gives 15,000 pages a month with no card, resetting monthly, which is the largest ongoing free allowance. Reducto includes 15,000 credits once. LlamaParse gives 10,000 credits a month. AWS Textract free usage lasts three months, and Azure F0 covers 500 pages a month but only processes the first two pages of each request.

What is the difference between a parser API and an extraction API?

A parser converts a document into text, markdown, or a layout tree, and you decide what the fields mean. An extraction API returns named fields for a known document type, such as invoice total or transaction amount. Parsers are cheaper and more general; extraction APIs do more of the work but only for the document types they support.

Why is per-page pricing misleading for document APIs?

Because most APIs bill per feature. AWS Textract charges separately for text, tables, forms, and queries, so a page that needs three of them costs three times. Google adds an hourly hosting fee for deployed custom processors, and Nanonets charges per pipeline step. Model your actual document flow rather than the headline rate.

Is there an API specifically for bank statements?

Yes. CapyParse exposes a REST API at /api/v1/conversions that returns normalized transactions as JSON, JSONL, CSV, or markdown, along with an MCP server for agent use. Mindee also publishes a bank statement model. Both do the transaction-level work that a general parser leaves to you.