Free OCR and Document AI APIs in 2026: Google Document AI, Azure, AWS Textract, Mistral OCR, Tesseract

OCR (optical character recognition) turns a scan or a photo of text into text you can search and copy. "Document AI" services go one step further: they also return tables, form fields and reading order. Free offers range from a few hundred pages a month to open-source engines you run yourself. Everything below was checked on each vendor's pricing or docs page on 2026-10-28.

Free OCR and document AI quotas side by side

Source: each vendor's pricing page or docs, linked in the table and under Sources, checked 2026-10-28.

Cloud quotas you can keep using

• Azure Document Intelligence has the most generous page count among the big clouds: the F0 tier lists "0 - 500 pages free per month". The catch is in the docs: the free tier "limits analyze response to only the first two pages in a request", with 4 MB per file and one request a second, so it suits a test or a form of a few pages, not a 100-page report.
• Google Document AI prices Enterprise Document OCR at $0.00 for the first 1,000 pages and $1.50 per 1,000 after that. Its setup page tells you to check that billing is enabled for the project, so the free pages sit behind a billing account.
• AWS Textract gives 1,000 pages a month of plain text detection, but only for the first three months, and 100 pages a month if you ask for tables, forms or layout. A synchronous call reads one page of a PDF; longer files go through the asynchronous API.

Free keys without a cloud account

• OCR.space is the easiest to try: register for a free key and send an image or a PDF of up to 3 pages. The free plan allows 25,000 requests a month, but only 1 MB per file and 500 requests a day per IP address.
• LlamaParse (LlamaCloud) gives 10,000 credits a month on the free plan. A page costs 1 credit in the fast tier, 3 in cost-effective and 10 in agentic, so the free allowance is about 10,000, 3,300 or 1,000 pages. It is aimed at feeding documents to an LLM.
• Mistral OCR is cheap rather than free: $2 per 1,000 pages, $1 in batch mode, according to its launch post. We could not confirm an OCR allowance in Mistral's free plan, so test with a small file before you plan on it. See the Mistral free plan.

Self-hosted: Tesseract, PaddleOCR, Docling

Tesseract is the classic open-source engine (Apache-2.0, 5.5.3 released 2026-07-24) with more than 100 languages. It returns text, not tables. PaddleOCR (Apache-2.0, 3.7.0 from 2026-06-11) handles many scripts and layouts, and can parse whole documents. Docling (MIT) converts PDFs and Office files to Markdown or JSON and can run OCR on scanned pages. All three cost only your machine, and the document never leaves it, which matters for contracts or medical records.

Which one to pick

• A few hundred pages for a test: Azure's 500 pages, or Google's first 1,000.
• Many small images, no cloud account: OCR.space.
• Tables and forms for a pipeline: Azure Layout, Textract tables (100 free pages a month) or LlamaParse. Check the result: scans of tables are where OCR fails most. Why is explained in why chatbots mangle tables in a PDF.
• Private documents or unlimited volume: Tesseract, PaddleOCR or Docling on your own machine.
• Non-Latin scripts and handwriting: AWS Textract reads handwriting in English only; try PaddleOCR or a cloud model on a sample first.

Steps

1. Pick one scan of a few pages and note whether it has tables.
2. For no-account testing, get a free key at ocr.space and run: curl -H "apikey: $OCR_SPACE_KEY" -F "file=@scan.png" -F "language=eng" https://api.ocr.space/parse/image
3. For a local run, install Tesseract and run tesseract scan.png out -l eng; the text lands in out.txt.
4. For a cloud test, create the resource (Azure F0 or a Google Document AI processor), send the same scan and compare the text with the Tesseract output.
5. Check the usage counter in the console before you send a large batch.

Limits and the catch

• Pages are counted differently. Azure counts every page analysed; Textract bills tables and forms apart from plain text; LlamaParse bills credits, not pages.
• Free allowances can end. Textract's free tier lasts three months; Google's and Azure's rules can change, so re-check the pricing page.
• Accuracy is not in the price list. A low-resolution scan, skewed text or a table can come back garbled on any service; check a few numbers by hand.
• Your documents are sent to the vendor. For contracts, ID cards or patient data read the vendor's data terms, or use a local engine.