Best OCR API for Indie Developers (2026 Prices)

Short answer: For most indie developers turning PDFs into clean markdown for RAG or automation, Mistral OCR is the best default at $4 per 1,000 pages, or $2 with batch. If you only need raw text, Azure Read, Google Enterprise Document OCR and Amazon Textract all cost $1.50 per 1,000 pages. If documents cannot leave your own servers, self-host Docling or PaddleOCR for free.

Every price and limit below comes from the vendor's own pricing page, docs or repo, checked on 13 Nov 2026. The advice in between is ours and is marked as such.

TL;DR#

• Best default for RAG and markdown: Mistral OCR 4 at "$4 per 1,000 pages", with "a 50% Batch-API discount, reducing the cost to $2 per 1,000 pages" (Mistral). The current model in the docs is OCR 4.1 at the same price (Mistral docs).
• Cheapest plain text: $1.50 per 1,000 pages on Azure Read, Google Enterprise Document OCR and Textract Detect Document Text.
• Tables and forms get expensive fast: Textract Forms is $50 per 1,000 pages (AWS).
• Best parsing pipeline with a free allowance: LlamaParse, 10K free credits a month for new users, then $1.50 per 1,000 credits (LlamaIndex).
• Free and private: Docling, PaddleOCR and Tesseract run on your own hardware.

Disclosure: ZeroRaised publishes this guide and is a launch site for bootstrapped products. ZeroRaised has no commercial relationship with any vendor named here and earns nothing if you sign up. Facts come from the linked sources as checked on 13 Nov 2026.

What "OCR API" means in 2026#

There are now three kinds of tool sold under the same label, and mixing them up is how indie developers overpay.

1. Classic OCR. Image in, text and coordinates out. Azure Read, Google Enterprise Document OCR, Textract Detect Document Text, Tesseract. Cheap, fast, and it knows nothing about what a table is.
2. Document parsers. PDF in, structured markdown or JSON out, with headings, tables and reading order intact. Mistral OCR, LlamaParse, Docling, PaddleOCR's structure models. This is what you want for retrieval-augmented generation, because chunking a table that has been flattened into one line of text ruins retrieval.
3. Field extractors. Invoice in, {"total": 142.50, "vendor": "..."} out. Textract Expense and Forms, Azure prebuilt models, Google Form Parser, or a vision LLM with a schema. Pair these with the techniques in our guide on getting structured output from an LLM.

Also check what counts as a billing unit. Most vendors bill per page, a single image counts as one page, and OCR.space bills per request instead, which matters for multi-page PDFs.

Comparison table#

Prices are per 1,000 pages at the entry tier, from each vendor's own page on 13 Nov 2026. "Not stated" means we could not find a claim on the vendor's own page today, not that the feature is missing.

Notes on the numbers:

• Textract prices are for US West (Oregon) on AWS's page, for the first 1M pages a month. Its handwriting support is English only (Textract quotas).
• Azure prices come from Microsoft's public retail prices API for East US, because the pricing page renders prices client-side. The F0 limit of "0 - 500 pages free per month" is from the pricing page. The handwriting language list is in the Read model language docs.
• Google handwriting detection is described in the Enterprise Document OCR docs. Google Cloud Vision is a separate, older text-detection API; its pricing page says the first 1,000 units a month are free.
• LlamaParse prices are our conversion: credits per page (1, 3, 10, 25) times $1.50 per 1,000 credits on pay-as-you-go. Parsed files are cached for 48 hours and "Re-parsing the same file within that window is free."
• Marker's code is Apache 2.0, but its README says the model weights are "free for research, personal use, and startups under $5M funding/revenue." Beyond that, you need a commercial licence.

We also looked at Reducto, Unstructured, Mindee and Nanonets. They are credible options, but we did not verify current per-page prices on their own pages today, so we have left them out of the table rather than guess.

The picks, one by one#

Mistral OCR: the default for RAG#

Mistral's launch post gives the price plainly and says the model "supports 170 languages across 10 language groups" (Mistral). Output is structured, with typed blocks for titles, tables, equations and signatures, which is exactly what you want to chunk for search. Mistral also publishes benchmark scores (85.20 on olmOCR-Bench, 93.07 on OmniDocBench) but calls them "directional" and says competitor scores are its own reproductions, so treat them as the vendor's claim.

Our read: $4 per 1,000 pages is low enough that a solo developer can stop thinking about it, and batch halves that for backfills. Its weak spot is that self-hosting is an enterprise conversation, not a download.

Azure, Google and Textract: cheap text, pricey structure#

All three charge $1.50 per 1,000 pages for plain OCR. The gap opens when you need structure: Azure's Layout model and prebuilt invoice or receipt models are $10, Google's Layout Parser $10 and Form Parser $30, and Textract charges $15 for tables and $50 for forms, adding up when you ask for both.

Our read: use the big clouds when you already live in that cloud, need a prebuilt receipt or ID model, or need handwriting. Azure has the clearest free tier for a side project (500 pages a month, no expiry stated on the page) and the broadest documented handwriting language list.

LlamaParse: when you want a pipeline, not a model#

LlamaParse lets you choose quality per job: Fast at 1 credit per page for clean digital PDFs, up to Agentic Plus at 25 credits for the ugly scanned stuff. The free allowance of 10K credits a month covers about 3,300 pages on the Cost-effective tier. The Pro plan is $500 a month with 400K credits included.

Vision LLMs: flexible, and cheaper than you think#

Google's docs say "Each document page is equivalent to 258 tokens" for PDFs up to 1,000 pages (Gemini document processing). That makes the input side almost free; the cost is the text you ask the model to write back. The upside is that one call can transcribe and extract fields into JSON. The downside is that LLMs can paraphrase or silently skip text, so check outputs on a sample before you trust them. Before sending customer documents anywhere, read our guide on whether AI APIs train on your data: Google's pricing table marks free-tier data as used to improve its products.

Self-hosted: Docling, PaddleOCR, Tesseract#

• Docling is MIT licensed, "started by the AI for knowledge team at IBM Research Zurich", reads PDF, DOCX, PPTX, XLSX, HTML and images, and supports "Local execution capabilities for sensitive data and air-gapped environments."
• PaddleOCR is Apache 2.0. Version 3.7.0 shipped on 11 June 2026, and its README describes PaddleOCR-VL as suited to "handwritten text and historical documents". The README reports 96.3% on OmniDocBench v1.6 for PaddleOCR-VL-1.6; that is the project's own reported result.
• Tesseract is Apache 2.0 and supports more than 100 languages out of the box. It is the lightest option, best on clean printed scans.

Our read: if you already run models on your own machine, Docling is the fastest way to get RAG-ready markdown with zero per-page cost.

Decision table by job#

Our editorial advice, based on the prices and features above.

Worked cost example: 20,000 pages a month#

Example (illustrative arithmetic, not a quote): an indie SaaS ingests 20,000 PDF pages a month of customer documents for a RAG feature. Prices are list prices from the sources above, ignoring tax and storage.

The Gemini line uses an assumption: 800 output tokens per page is our guess for a dense page of markdown. Your real number depends on your documents and prompt, so measure 50 pages before you budget. Input is $0.25 and output $1.50 per 1M tokens on the paid tier (Gemini pricing).

Our read: the gap between $30 and $1,300 a month is not about OCR quality, it is about paying for structure you may not need. Start with a document parser (Mistral or Docling) and only pay for forms extraction on the document types that actually need it. The same habits cut any AI bill: batch what can wait, cache what repeats, route easy inputs to cheap models.

How to choose in an afternoon#

1. Collect 30 to 50 real pages, including your worst scans, tables and any handwriting.
2. Run them through one cheap text OCR (Azure Read or Google), one parser (Mistral or Docling) and, if you need fields, one vision LLM with a schema.
3. Diff outputs against a hand-checked version. Count broken tables and missing lines, not vibes.
4. Price the winner at your expected monthly volume using the table above, then at 10 times that volume.
5. Check data terms before production. If privacy is the deciding factor, self-host.

Frequently asked questions#

What is the cheapest OCR API in 2026?#

For plain text, Azure Read, Google Enterprise Document OCR and Amazon Textract Detect Document Text all list $1.50 per 1,000 pages, and Azure and Google drop to $0.60 at high volume. OCR.space has a free plan with 25,000 requests a month, but free PDFs are limited to 3 pages. Self-hosted Tesseract, PaddleOCR or Docling cost nothing in licence fees.

Is Mistral OCR worth it over Textract or Azure?#

For RAG and document-to-markdown work, usually yes. Mistral OCR 4 is $4 per 1,000 pages ($2 in batch) and returns structured output including tables. Textract and Azure are cheaper for plain text but charge $10 to $50 per 1,000 pages once you need tables, forms or prebuilt models.

Which OCR API handles handwriting?#

Azure's Read model documents handwriting in several languages, including English, French, German, Spanish, Japanese and Chinese Simplified. Google Enterprise Document OCR detects handwriting by default. Textract recognises handwriting in English only. OCR.space supports handwriting on its Engine 3. PaddleOCR-VL is described as suitable for handwritten text.

Can I use a vision LLM like Gemini as an OCR API?#

Yes. Google's docs say each PDF page counts as 258 tokens. On gemini-3.1-flash-lite at $0.25 per 1M input tokens and $1.50 per 1M output tokens, our labelled example of 800 output tokens a page works out to about $1.26 per 1,000 pages. Note that free-tier Gemini data is used to improve Google's products.

What is the best free OCR to self-host?#

Docling (MIT, started at IBM Research) is the easiest route from PDFs to markdown with table structure. PaddleOCR (Apache 2.0) is strong on tables and handwriting with its VL models. Tesseract (Apache 2.0) is the lightweight classic for clean printed text. Check Marker's weights licence before commercial use.

Ship the document feature, then launch it#

If OCR is the engine behind your product, the next step is getting it in front of users. The product launch checklist for bootstrapped makers covers launch day. When you are ready, launch on ZeroRaised for free, and join the newsletter for more guides on building without investors.

Sources#

• Mistral OCR 4 announcement, last checked 13 Nov 2026
• Mistral OCR 4.1 model docs, last checked 13 Nov 2026
• Amazon Textract pricing, last checked 13 Nov 2026
• Amazon Textract quotas (handwriting), last checked 13 Nov 2026
• Azure AI Document Intelligence pricing, last checked 13 Nov 2026
• Azure retail prices API, Document Intelligence, East US, last checked 13 Nov 2026
• Azure Read and Layout language support, last checked 13 Nov 2026
• Google Document AI pricing, last checked 13 Nov 2026
• Google Enterprise Document OCR docs, last checked 13 Nov 2026
• Google Cloud Vision pricing, last checked 13 Nov 2026
• LlamaIndex pricing, last checked 13 Nov 2026
• LlamaParse pricing docs, last checked 13 Nov 2026
• Gemini API pricing, last checked 13 Nov 2026
• Gemini document processing docs, last checked 13 Nov 2026
• OCR.space OCR API, last checked 13 Nov 2026
• Docling on GitHub, last checked 13 Nov 2026
• PaddleOCR on GitHub, last checked 13 Nov 2026
• Tesseract on GitHub, last checked 13 Nov 2026
• Marker on GitHub, last checked 13 Nov 2026