What is the difference between OCR, layout parsing and structured extraction?

OCR turns pixels into a flat string of text. Layout parsing keeps the document's structure — headings, reading order, lists and tables — usually emitted as Markdown or HTML, which is what retrieval-augmented generation pipelines want. Structured extraction goes further and returns named fields against a schema, such as invoice_number or total. They are priced very differently: at list prices in 2026, plain OCR runs about $1.50 per 1,000 pages from the hyperscalers, layout parsing about $10 per 1,000, and schema extraction $30 to $50 per 1,000.

What is the cheapest document parsing API in 2026?

For plain text and layout-aware Markdown at volume, flat per-call APIs are cheapest on published list prices — Carterstack's Smart OCR API works out to roughly $0.53 per 1,000 pages on its $79/month plan. AWS Textract, Azure AI Document Intelligence Read and Google Enterprise Document OCR are the cheapest hyperscaler options at about $1.50 per 1,000 pages. Self-hosting an open-source parser such as Docling or Marker has no per-page licence cost at all, but you pay for the GPU and the operational work instead. Enterprise volumes are almost always negotiated below list.

Should I self-host an open-source parser instead of paying for an API?

If you are parsing millions of pages and already run GPU infrastructure, self-hosting Docling, Marker or MinerU is usually the cheapest option, because the marginal cost per page approaches zero. Below that scale the maths rarely works: a managed API costs less than the engineering time you would spend on model updates, queueing, retries and monitoring. The crossover point in practice is somewhere in the hundreds of thousands of pages per month, and it moves depending on how much your team's time is worth.

← Blog

Best document parsing APIs for developers in 2026

July 29, 2026 · 7 min read · OCR & Documents

"Document parsing API" is doing a lot of work as a phrase. It's used to sell three genuinely different products at three wildly different prices, and picking the wrong one is how teams end up paying $30 per 1,000 pages for something they could have bought for $1.50.

This is a pricing-first comparison. Every number below is a published list price as of July 2026, linked to its source at the bottom. Nobody here is paying us, and where a competitor is the better choice, it says so.

First, work out which of the three jobs you have

Almost every mis-purchase starts here. The three tiers, cheapest first:

Plain OCR. Pixels in, a flat string of text out. Fine for search indexing, moderation queues, or anything where you just need the words. Roughly $1.50 per 1,000 pages at the hyperscalers.
Layout parsing. Text plus structure — headings, reading order, lists, tables — usually as Markdown or HTML. This is what RAG pipelines actually want, because a table flattened into prose retrieves terribly. Roughly $10 per 1,000 pages.
Structured extraction. Named fields against a schema: invoice_number, total, candidate_email. Roughly $30–$50 per 1,000 pages, or a per-document price from a specialist.

If you're building RAG, you want tier two and you are probably being sold tier three. If you're building an expense or ATS feature, you want tier three and plain OCR will waste a month of your life in regex.

Document parsing API pricing, compared

List prices, as of July 2026. Normalised per 1,000 pages where the vendor publishes enough to do so honestly.

Provider	List price (2026)	Tier	Free tier	Where it genuinely wins
AWS Textract	$1.50/1k OCR · $15/1k tables · $50/1k forms	All three	1,000 pages/mo, 3 months	Scale, VPC, data residency. Forms/key-value pairs are excellent.
Azure AI Document Intelligence	$1.50/1k Read · $10/1k Layout · $10/1k prebuilt · $30/1k custom	All three	500 pages/mo (2 pages per request)	Cleanest tier ladder; strong table extraction; custom models.
Google Document AI	$1.50/1k OCR (→$0.60 over 5M) · $10/1k Layout Parser · $30/1k Form Parser	All three	Trial credits	Layout Parser chunks for RAG in the same call. Cheapest at huge scale.
Mistral Document AI	~$4/1k pages OCR · ~$5/1k Document AI	OCR + layout	Trial credits	Very good Markdown out of the box; EU hosting; simple API.
LlamaParse (LlamaCloud)	1,000 credits = $1.25; 1 credit/page basic, far more for agentic modes. Starter $50/mo	Layout + extraction	10,000 credits/mo	130+ input formats, RAG-native output, SOC 2 / HIPAA.
Unstructured	$0.03/page ($30/1k); bill caps at $3,000/mo	Layout + chunking	15,000 pages/mo	Enormous free tier; connectors to 30+ sources; batteries-included ingestion.
Reducto	$0.015/credit after free credits (~$15/1k basic parse); agentic OCR 2×	Agentic layout + extraction	15,000 credits	Best-in-class on nasty tables and financial PDFs. ZDR, BAA, VPC on higher tiers.
Adobe PDF Extract	Quote only above the free tier	Layout	500 transactions/mo	Native-PDF structure fidelity from the people who invented the format.
Docling / Marker (open source)	$0 licence — you pay GPU + ops	Layout	Free forever	Zero marginal cost at millions of pages. Fully private.
LLM vision (GPT/Claude/Gemini)	Per token — varies with page density	Anything, unreliably	Varies	Odd one-off documents and schema-flexible extraction.
Carterstack (Smart OCR)	$9.99/mo → 5,000 calls (~$2.00/1k); $24.99 → 25,000 (~$1.00/1k); $79 → 150,000 (~$0.53/1k)	OCR + layout Markdown	100 req/mo	Lowest per-page list price for layout-aware Markdown. Flat, predictable.

The math at 100,000 pages a month

Per-page rates are hard to feel. Here's one workload — 100,000 pages, layout-aware Markdown wanted — at each vendor's published rate:

Provider	Monthly cost at 100,000 pages
Docling / Marker, self-hosted	GPU + engineering time only
Carterstack Smart OCR (Mega, 150k quota)	$79
LlamaParse, basic 1-credit mode	~$125 ($50 Starter + 60k credits)
Mistral Document AI	~$400–500
Azure Layout / Google Layout Parser	~$1,000
Reducto (basic parse, after free credits)	~$1,275+ — get a quote at this volume
Unstructured	~$2,550 (15k free + 85k × $0.03)

Two honest caveats. First, this table is not apples-to-apples: Reducto and Unstructured are doing considerably more per page than raw OCR — chunking, table reconstruction, agentic re-reads — and some of that you'd otherwise build yourself. Second, nobody at real volume pays list. Every vendor above except the self-serve tiers will discount, sometimes steeply.

Where the competition is genuinely better than us

If this only listed our own advantages it would be an ad, not a comparison. So:

Unstructured gives you 15,000 free pages every month. Ours is 100 requests. For prototyping and for a lot of small production workloads, theirs is simply free and ours isn't.
LlamaParse ingests 130+ formats — PDF, DOCX, XLSX, PPTX — natively. We take images. If your inputs are a mixed bag of office files, that difference will cost you a day of glue code you don't need to write.
Reducto is the one to beat on genuinely hostile documents: merged-cell financial tables, multi-column research papers, forms with handwriting over printed rules. If accuracy on those pages is the product, pay for it.
AWS Textract and Azure give you regional endpoints, VPC isolation and data-residency guarantees. If your legal team has opinions about where a document is processed, this isn't a close call.
Google's Layout Parser returns RAG-ready chunks in the same call, which saves a whole step in your ingestion pipeline.
Docling and Marker cost nothing per page, run entirely on your own machines, and are now genuinely good. At millions of pages, self-hosting wins on price and nothing we can do changes that.

Where Carterstack Smart OCR actually competes

One narrow claim: layout-aware Markdown at plain-OCR prices, with a bill that doesn't move.

Most vendors charge roughly 7× more for structure than for raw text — $1.50 for Read, $10 for Layout. Smart OCR returns either, at the same flat per-call price, which at the $79 tier is about $0.53 per 1,000 pages. Set "format": "markdown" and headings, bullet and numbered lists, and tables come back as real Markdown. Printed or handwritten, and it auto-detects language.

curl -X POST \
  "https://smart-ocr-image-document-to-text.p.rapidapi.com/v1/ocr" \
  -H "Content-Type: application/json" \
  -H "X-RapidAPI-Key: YOUR_KEY" \
  -H "X-RapidAPI-Host: smart-ocr-image-document-to-text.p.rapidapi.com" \
  -d '{ "image_url": "https://example.com/page.jpg", "format": "markdown" }'

Back comes text, format, char_count and word_count. That's the whole response — no async polling, no job IDs.

Why we can price this way: all four Carterstack APIs run inference on hardware we own. There's no upstream token meter, so there's nothing to pass through to you — pricing is flat per call, and a dense 900-word page costs exactly the same as a receipt. No per-call token games.

What we don't do, plainly

Smart OCR takes one image per call. There's no native PDF ingestion — a 10-page PDF is 10 calls, and you rasterize the pages yourself. Factor that into the comparison above; on a per-page basis the price holds, but the plumbing is on you. We also have no schema-extraction endpoint, no chunking, no review UI and no SOC 2 report. For field-level work on receipts, invoices or resumes we have purpose-built APIs instead — but for arbitrary schemas on arbitrary documents, buy from someone above.

So which should you pick?

Building RAG over messy real-world PDFs? Google Layout Parser or LlamaParse. Reducto if the tables are brutal and the budget allows.
Already deep in AWS or Azure with residency requirements? Textract or Azure Document Intelligence. $1.50/1k for Read is genuinely good.
Millions of pages and a GPU team? Self-host Docling or Marker. The API bill stops being worth it.
Prototyping, or under 15,000 pages a month? Unstructured's free tier, honestly.
High volume of image-based pages, want text or Markdown, want a bill you can predict? That's the gap we built for.

Whatever you shortlist, benchmark it on your documents — the photographed contract, the scan that's 3° off square, the table that spans a page break. Vendor demo files tell you nothing. Two hours on fifty real pages will tell you more than any comparison table, including this one.

Run it on your worst page

Smart OCR has a free tier — no credit card to start. All four Carterstack APIs run on owned hardware, so pricing is flat per call: no per-token games, no surprise bill on dense pages.

See the API & grab a free key →

Not affiliated with, sponsored by, or endorsed by any vendor listed. All pricing is publicly listed pricing as of July 2026 and is summarised in good faith — vendors change rates, credit systems vary by document complexity, and enterprise pricing is negotiated. Verify before you budget: AWS Textract, Azure AI Document Intelligence, Google Document AI, Mistral, LlamaParse, Unstructured, Reducto, Adobe. Carterstack runs its APIs on its own hardware.