2kw.ai
Sign inBook a 15-min demo
Compare · Document extraction APIs

The best document extraction APIs in 2026

Six APIs that turn PDFs, scans, spreadsheets and emails into structured data — compared on capability, free allowance and published price, with every figure read off the vendor's own page.

The short answer

A document extraction API turns an unstructured file into structured data. Two shapes exist: OCR-style APIs that return text and layout for you to interpret, and schema-driven APIs where you declare the fields you want and get validated JSON back. Pick the first when you do not know what a document contains, the second when you do.

The options at a glance

OptionWhat it isFree allowancePublished price
2kw.aiSchema-driven extraction: you define a JSON Schema and get validated structured output, with a grounding score linking each value back to the source document. 17 input formats including PDF, DOCX, XLSX, PPTX, EML, ZIP, images and the engineering formats DXF, DWG, STEP and GEO; output as JSON, Markdown, plain text or HTML. Audio transcription with SRT and VTT on the same account.7-day free trial on every paid plan. No permanently free tier.€99 / €299 / €499 per month, including 200 / 500 / 1,000 extraction credits and 1,000 / 5,000 / 10,000 conversion credits. Extraction is unmetered with BYOK on Scale and Enterprise. The credit-to-page mapping is not published — see the note below, and hold us to it.
Amazon TextractAWS OCR and document analysis: text detection, plus table and form extraction through the Analyze Document API.Three months: 1,000 pages per month for text detection, 100 pages per month for forms, tables and layout.US West (Oregon), USD. Detect Document Text $1.50 per 1,000 pages for the first 1M pages, $0.60 beyond. Analyze Document — Tables $15.00 per 1,000 pages, $10.00 beyond 1M. Forms $50.00 per 1,000 pages, $40.00 beyond 1M.
Azure AI Document IntelligenceMicrosoft's document model family: Read, Layout, prebuilt models for common document types, and custom classification and extraction models.0–500 pages free per month.Billed per 1,000 pages by model type. The pricing page renders its rates dynamically and states that prices are estimates varying by region, currency and agreement, so no per-page figure is quoted here — use the Azure pricing calculator for your region.
LlamaParse“An agentic document parser built for LLM pipelines — layout-aware OCR that turns PDFs, scans, tables, and charts into clean markdown, text, or JSON.” 130+ file types.Free plan at $0/month, including 10K credits.Starter $50/month (40K credits), Pro $500/month (400K credits), Enterprise custom. 1,000 credits = $1.25. Basic parsing costs “as low as 1 credit”; layout-aware agentic parsing with LLMs or VLMs costs more per page.
Reducto“The complete agentic document platform” — parsing, extraction, editing and splitting, with OCR, table recognition and form parsing across 30+ file types.First 15K credits free.$0.015 per credit after the first 15K. Growth and Enterprise tiers priced on request.
UnstructuredTurns documents into LLM-ready data: extraction, partitioning, chunking, enrichment and embedding generation across 50+ document and image file types. An open-source library exists alongside the hosted platform.10,000 pages free to start, no card required, all features included.$0.015 per page pay-as-you-go after the free pages. Business tier custom.

Every cell was verified against the linked vendor page on 31 August 2026. Those pages are the authority — pricing and features change without notice. 2kw.ai is one of the options listed and has an obvious interest in the outcome; its row is marked so you can weigh it accordingly.

Listed alphabetically. The order is not a ranking.

The prices in that table do not compare, and that is the real problem

Every column but the last is straightforward. The last one is a trap, because these vendors do not sell the same unit.

Amazon Textract charges per page and per API. The same page costs $1.50 per thousand through Detect Document Text and $50.00 per thousand through Analyze Document Forms — a factor of over thirty, decided by which call you make, not by the document. Any comparison that quotes a single Textract number has already chosen an answer for you.

LlamaParse and Reducto charge per credit, and a credit is not a page. LlamaParse states basic parsing costs “as low as 1 credit” while layout-aware agentic parsing with LLMs or VLMs costs more, so the per-page cost moves with the parsing mode you select. At 1,000 credits for $1.25, a one-credit page is $0.00125 and a ten-credit page is $0.0125 — an order of magnitude apart on the same price list.

Unstructured is the one clean per-page number in the table at $0.015 per page. Azure publishes per-1,000-page pricing but renders it dynamically and calls it an estimate that varies by region and agreement. And 2kw.ai bundles credits into a subscription without publishing what a credit is worth in pages, which makes our own row the hardest one here to price-compare. That is a gap in our documentation and it is being fixed.

The only comparison that survives contact with reality is your own: take a hundred representative documents, run them through two or three candidates, and price the actual month. Every vendor on this page including us offers a free allowance large enough to do that.

Which one fits your situation

You are already in AWS or Azure
Textract and Azure AI Document Intelligence keep the data inside the cloud account you already have, with the IAM, networking and billing you already run. That is worth a lot, and it is usually the deciding factor when it applies. Both are OCR-and-layout products: you get text, tables and form key-value pairs, and the interpretation is yours.
You are building a RAG pipeline
LlamaParse and Unstructured are built for this. Unstructured goes furthest — partitioning, chunking, enrichment and embeddings in one pass — and its open-source library means you can start without a vendor at all. LlamaParse covers 130+ file types and outputs clean markdown, which is usually what a retrieval pipeline wants.
You know exactly which fields you need
Schema-driven extraction is a different job from OCR. Declare the fields and types once and get validated JSON; anything that does not match surfaces at extraction time rather than three services downstream. Reducto and 2kw.ai both work this way, and both handle documents too large for a single pass.
You have to prove where a number came from
In regulated or financial work, an extracted value is only useful if it can be traced. 2kw.ai returns a grounding score per value linking it back to the source document, which is what lets low-confidence fields go to human review while the rest clear automatically. Check which of the others provide this before assuming it is standard — it is not.

Common questions

What is a document extraction API?

A document extraction API takes an unstructured file — a PDF, a scan, a spreadsheet, an email — and returns structured data from it. The two common shapes are OCR-style APIs that return text and layout for you to interpret, and schema-driven APIs where you declare the fields you want and receive validated JSON. Which shape you need depends on whether you know in advance what you are looking for.

Why can't I just compare the per-page prices?

Because the vendors do not sell the same unit. Amazon Textract charges per page and per API, so a page run through Forms costs over thirty times what the same page costs through text detection. LlamaParse and Reducto charge per credit, and a credit is not a page — LlamaParse states basic parsing starts at one credit while agentic parsing costs more. Subscription products bundle an allowance. Price a realistic month of your own documents against each model rather than comparing headline rates.

What does 2kw.ai's extraction credit correspond to in pages?

That mapping is not currently published, which makes 2kw.ai harder to price-compare against the per-page vendors in this table than it should be. It is a real gap in our own documentation rather than a subtlety, and it is being fixed. Until it is, ask us for the mapping for your document mix, or run the 7-day trial on your own files and measure it.

Do I need OCR, or do I need schema-driven extraction?

If you do not know in advance which fields a document contains, you need OCR and layout — Textract, Azure Read and Layout, or LlamaParse. If you do know, schema-driven extraction skips the interpretation step: you declare the fields and their types once and get validated JSON, and anything that does not match the schema surfaces immediately instead of downstream. Many pipelines use both, OCR first and schema second.

What is a grounding score and why does it matter?

A grounding score links an extracted value back to the place in the source document it came from. Without it, a wrong value and a right value look identical in the output, and verification means a human re-reading the original. With it, low-confidence fields can be routed to review and the rest can be trusted automatically. 2kw.ai returns one per extracted value; check which of the others do before assuming it is standard.

Why is Google Document AI not in this table?

Because its pricing page did not yield figures that could be quoted verbatim on 31 August 2026, and this page does not publish numbers it has not verified. Google Document AI is a serious option and its absence is a limitation of this comparison, not a judgement on the product.

How this comparison was made

2kw.ai is our product and appears in the table, so read this as a vendor-published comparison. Every competitor figure was read off that vendor's own page on 31 August 2026and is quoted or closely paraphrased. Nothing was estimated or taken from a third-party article. Where a page did not state something — Azure's per-page rates, Google Document AI's pricing altogether — the page says so instead of filling the gap.

Prices and capabilities change without notice. If an entry is wrong or has gone stale, tell us and it gets corrected.

More from the platform
Document extraction Schema-driven structured data extraction from 17 document formats, with grounding scores.Best LLM gateways Nine gateways compared on hosting model, OpenAI compatibility and published price.Best EU-hosted LLM gateways The options that keep LLM traffic inside the European Union.Cost tracking Per-token and per-field cost tracking across providers.

Define a schema. Get validated JSON with the source to back it.

Start 7-day free trialBook a 15-min demo