Every RAG demo looks great until you feed it a real PDF. A scanned contract with a signature block. A bank statement where the numbers sit in a borderless table. A two-column research paper with footnotes that wrap. The retrieval, the embeddings, the fancy reranker — none of it matters if the text that went in was mangled at the door. Document extraction is the layer most teams pick last and regret first.
So here’s a look at the options I’d actually consider in 2026: the cloud OCR veterans (AWS Textract, Azure Document Intelligence, Google Document AI), the layout-aware parsers (Unstructured, LlamaParse, Docling, Marker), and the newer VLM-based “document understanding” crowd led by Mistral OCR 3. They don’t all solve the same problem, and the price spread between them is wild — from free to roughly $60 per 1,000 pages depending on what you ask for.
Three ways to turn a PDF into something an LLM can use
Classic OCR. AWS Textract, ABBYY, Google’s OCR processor, Azure’s Read model. These are pixel-to-text engines, refined over a decade, very good at “what characters are on this page.” Cheap and fast. What they don’t do: tell you that this block is a table header and that one is a footnote. You get text; structure is your problem.
Layout-aware parsers. Unstructured, LlamaParse, Docling, Marker, MinerU. These run detection models to find titles, paragraphs, tables, lists, and figures, then emit Markdown or JSON that preserves the hierarchy. This is the sweet spot for most RAG pipelines — chunk on real section boundaries instead of arbitrary character counts and your retrieval gets noticeably better, with no other change.
VLM-based document understanding. Mistral OCR 3, GPT-5 vision, Gemini 3. Feed the page image to a vision-language model and ask it to read like a human — reconstruct reading order, handle handwriting, pull a strict JSON schema straight out of an invoice. The most flexible and the slowest, and the one that hallucinates if you don’t constrain it with a schema and a validation step.
The line between the last two has gotten blurry. LlamaParse’s “agentic” mode is basically a VLM under the hood. Mistral OCR 3 is sold as OCR but behaves like a parser. Don’t get hung up on the category names — get hung up on what the output looks like on your documents.
Accuracy: who’s actually winning
Benchmarks here are messy, because “accuracy” depends entirely on what you’re parsing. On clean digital PDFs almost everything scores in the 90s and the differences are noise. The gaps open up on the hard stuff: complex tables, multi-column layouts, skewed scans, handwriting.
From the public benchmarks I trust most as of early 2026 — including the ParseBench evaluation that’s been making the rounds — the agentic parsers cluster at the top. LlamaParse’s agentic mode lands around 72% on the hardest document mix, Gemini 3 Flash around 71%, Reducto around 68%. Mistral pitched OCR 3 with a 74% win rate against competing products on forms, scanned docs, complex tables, and handwriting. That’s a vendor number — treat it accordingly — but it tracks with what I’ve seen: Mistral is genuinely strong on tables and handwriting.
Unstructured, meanwhile, publishes its own benchmarks showing it leading on parsing quality across formats, which is also a vendor number. You see the pattern. Everyone wins their own benchmark, so the only score that counts is the one you run yourself.
Textract, Azure Read, and Google’s base OCR aren’t in this race and aren’t trying to be — they’re the floor, not the ceiling. If your documents are mostly clean and mostly text, the floor is fine, and it’s a lot cheaper.
My rough mental model: for born-digital PDFs, use the cheapest thing that emits decent Markdown. For tables and forms, test Mistral OCR 3 and Reducto head to head. For handwriting and genuinely degraded scans you’re in VLM territory whether you like the price or not. And always — always — run your own eval set, because the document that breaks your pipeline is never the one in anyone’s benchmark.
What it costs per 1,000 pages
Here’s where the decision usually gets made, because the spread is enormous. Prices below are roughly mid-2026; check the vendors before you commit, because this category re-prices constantly.
- Open-source, self-hosted (Docling, Marker, MinerU): $0 in license; you pay for compute and engineering time. On a CPU box, Marker does roughly 1–2 seconds per simple page and 3–4 seconds on complex ones; Docling is slower but more accurate on complex tables and document structure. Both have first-class loaders for LangChain and LlamaIndex. Realistically a few cents per 1,000 pages in cloud compute.
- Mistral OCR 3: $2 per 1,000 pages on the API, $1 per 1,000 with the batch API. Self-hosting available for data-residency cases, and it powers the drag-and-drop Document AI tool in Mistral’s studio. This is the price-to-quality standout of 2026.
- Reducto: credit-based — 1,000 credits costs $1 in North America; standard parsing is 1–2 credits per page, agentic parsing (the “editor” pass that corrects OCR errors) is 2–4. So call it $1–$2 per 1,000 pages for a normal parse, more for the agentic mode. Enterprise plans go lower at volume.
- LlamaParse: 10,000 free credits a month, then credit-based. “Parse without AI” is 1 credit per page, the cost-effective LLM mode is about 3 credits per page, and the agentic mode with a state-of-the-art model can run up to ~90 credits per page. Credits map to dollars on your plan; the takeaway is that the cheap mode is genuinely cheap and the premium mode very much is not.
- AWS Textract: basic Detect Document Text is $1.50 per 1,000 pages, dropping to $0.60 past a million. But the moment you want tables it’s $15 per 1,000, forms $50 per 1,000, the Analyze Expense API for receipts and invoices $8 per 1,000. Textract’s pricing punishes you for needing structure.
- Azure Document Intelligence: the Read model is $1.50 per 1,000 pages, prebuilt models (invoice, receipt, ID) about $10 per 1,000, custom extraction roughly $30 per 1,000. Annual volume commitments cut the effective rate hard if you’re processing millions of pages a month.
- Google Document AI: the OCR processor is around $1.50 per 1,000 pages, Form Parser much more, custom extractors about $0.10 per page plus training fees — and don’t miss the roughly $36/month per deployed processor version hosting fee, which stings if you run several custom processors at low volume. New accounts get $300 in credit, which is plenty for testing.
- Unstructured: free open-source library, plus a serverless API and an enterprise Platform (SOC 2 Type II, HIPAA, in-VPC deployment) priced per page on paid tiers. If you already self-host the OSS version, the API mostly buys you not maintaining it. Thirty-plus input formats and several chunking strategies, which is more than most.
Two things that bite people. First, “OCR” pricing and “extraction” pricing are different products at AWS, Azure, and Google — the headline $1.50 is text-only; structured data costs 5 to 30 times that. Second, the cloud providers layer in hosting and deployment fees that never show up in a per-page calculator. Mistral, Reducto, and LlamaParse are refreshingly close to what-you-see-is-what-you-pay.
And one tier above all of this sits the legacy enterprise IDP world — ABBYY, Rossum, Hyperscience — sold on annual contracts with implementation services, aimed at insurers and banks digitizing decades of paper. If you’re not in that world you don’t need to think about them, and if you are, you already have a sales rep.
Pick by scenario
Side project or low volume. Start with Docling or Marker self-hosted (free), or LlamaParse on its 10,000 free credits. Don’t pay anyone until you hit a wall. If you do hit one, Mistral OCR 3 at $1–$2 per 1,000 pages is the obvious next step.
High-volume ingestion, millions of pages. This is where per-page price compounds into real money. Mistral OCR 3’s batch API at $1 per 1,000 pages is hard to beat on quality per dollar. If you need on-prem or zero data retention for compliance, Reducto and Unstructured both deploy in your VPC and have the SOC 2 and HIPAA paperwork ready. The hyperscalers make sense here mostly if you’re already deep in one and have negotiated volume pricing.
Already an Azure, AWS, or GCP shop. Use the native service for the boring 80% — Azure Read, Textract Detect, Google OCR — and reach for a specialist only when accuracy on a specific document type isn’t good enough. The integration savings are real: same IAM, same VPC, same bill, no new vendor security review. Just go in knowing the structured-extraction tiers are pricey.
Schema-strict extraction — invoices, forms, statements. You want JSON that matches your schema, not Markdown you then parse again. Azure’s prebuilt invoice and receipt models, AWS’s Analyze Expense, and Mistral’s structured-output mode all do this. For weirder or industry-specific forms, a VLM with a tight JSON schema and a validation step beats fighting a rules-based template. Reducto’s Extract endpoint is built for exactly this and worth a look when accuracy can’t slip.
Building on LlamaIndex or LangChain. LlamaParse drops straight into LlamaIndex with bounding boxes and citations; Docling and Marker have first-class loaders for both frameworks. Path of least resistance, and you can swap the parser later without touching your retrieval code.
Wiring it in without shooting yourself in the foot
A few things I wish someone had told me earlier.
Parse to Markdown when you can, not plain text. Markdown keeps headings and tables legible to the model and gives your chunker something real to split on. Chunk on section boundaries — the whole reason you paid for a layout-aware parser — not a fixed token window that slices a table in half.
Keep bounding boxes and page numbers in your metadata. The day a user asks “where did this number come from,” you’ll want to point at a region on a page, not shrug. Citations in RAG are only as good as the provenance you carried through from extraction.
For tables, decide early whether you want them as Markdown, as HTML, or as structured rows. LLMs read Markdown tables fine for Q&A but struggle to do arithmetic across them; if you need the numbers, pull them as structured data and let code do the math.
And eval extraction on its own, before you blame the model. Build a small set of your nastiest documents, define what “correct” output looks like, and diff every parser against it. This is the single highest-leverage hour you’ll spend on a document pipeline, and almost nobody does it until something’s on fire. If you already run an LLM eval or observability setup — Langfuse, LangSmith, Braintrust — point it at the extraction step too; it’s just another stage that can silently regress.
What I’d reach for today
Starting a document-heavy project this week: Docling or the LlamaParse free tier to prototype, Mistral OCR 3 the moment I need better tables or handwriting and want predictable pricing, Reducto or Unstructured if compliance and on-prem are hard requirements, and whichever hyperscaler I’m already on for the simple bulk text. The cloud OCR veterans aren’t obsolete — they’re just no longer the default, and they make you pay a structure tax the newer players don’t.
Grab five of your worst PDFs and run them through three of these tomorrow. The winner is almost never the one with the best benchmark blog post.