Best OCR API in 2026: 13 APIs compared on price and real tests
13 OCR APIs compared: price per 1,000 pages, tables, handwriting and forms, with results from our own tests. Plus free OCR APIs and the best OCR SDK options.
Updated
TL;DR
- For most documents, Mistral OCR is the best default: fast, $25 per 1,000 pages, and it kept table structure in all our tests.
- For handwriting, Jina OCR (98.9% word accuracy) and Mistral OCR (96.8%) led our test; Google Document AI scored 76.6%.
- For the lowest cost, DeepSeek-OCR and Jina OCR cost $1 per 1,000 pages; Tesseract is free.
- For forms and IDs, AWS Textract and Azure Document Intelligence return key-value pairs, boxes and confidence scores.
An OCR API takes an image or PDF and returns the text, usually as plain text, Markdown or JSON. There are dozens. This page compares the 13 we run in production behind one API, using the price we charge per 1,000 pages and the results of tests we published with the files, so you can check them.
The short list
| Use case | Best pick | Price per 1,000 pages | Why |
|---|---|---|---|
| General documents to Markdown | Mistral OCR | $25 | Kept tables in every test, about 0.9 s per page |
| Handwriting | Jina OCR | $1 | 98.9% word accuracy on our handwritten note |
| Lowest cost | DeepSeek-OCR | $1 | Cheap bulk conversion with tables and formulas |
| Forms, key-value pairs, IDs | AWS Textract | $25 | Returns fields with boxes and confidence scores |
| Enterprise forms, many languages | Azure Document Intelligence | $60 | Forms, handwriting, boxes and confidence in one call |
| Chinese, Japanese, Korean | PaddleOCR-VL | $6 | Near the top of OmniDocBench, built for CJK documents |
| Free | Tesseract | Free | Clean printed text only; fails on handwriting |
All 13 OCR APIs compared
Prices are what you pay on OCRRouter, checked November 17, 2026. "Speed" is our typical latency per page. Yes means the API returns that kind of output; it says nothing about how well. Handwriting is covered by the test below.
| API | Type | Price / 1,000 pages | Speed | Tables | Forms | Boxes |
|---|---|---|---|---|---|---|
| DeepSeek-OCR | Open model, hosted | $1 | 1.1 s | Yes | No | Yes |
| Jina OCR | Open model, hosted | $1 | 1.5 s | Yes | No | No |
| Qianfan OCR | OCR API | $2 | 2.0 s | Yes | No | No |
| PaddleOCR-VL | Open model, hosted | $6 | 1.3 s | Yes | No | Yes |
| GLM-OCR | Open model, hosted | $6 | 1.3 s | Yes | No | Yes |
| Google Document AI | OCR API | $10 | 1.4 s | Yes | Yes | Yes |
| GPT | Vision model | $11 | 2.4 s | Yes | No | No |
| Gemini Flash | Vision model | $15 | 1.6 s | Yes | No | No |
| Mistral OCR | OCR API | $25 | 0.9 s | Yes | No | Yes |
| AWS Textract | OCR API | $25 | 1.2 s | Yes | Yes | Yes |
| Claude | Vision model | $25 | 2.6 s | Yes | No | No |
| Azure Document Intelligence | OCR API | $60 | 1.5 s | Yes | Yes | Yes |
| Tesseract | Open engine, hosted | Free | 0.6 s | No | No | No |
Three kinds of API sit in this table, and they behave differently:
- Dedicated OCR APIs (Textract, Azure, Google Document AI, Mistral OCR) are built for documents. They are predictable and return structure such as boxes, fields and confidence.
- Open document models (PaddleOCR-VL, GLM-OCR, DeepSeek-OCR, Jina OCR) are newer vision-language models trained for OCR. They are cheap and score well on public benchmarks.
- General vision models (Gemini, GPT, Claude) are prompted to transcribe. They handle odd layouts and instructions well, cost more per page and can occasionally rephrase text instead of copying it.
What our tests show
We ran the same files through the APIs above on November 14 and 15, 2026. Each test has its own write-up with the file.
Tables. On a printed invoice (details), Mistral OCR, GLM-OCR, PaddleOCR-VL, Jina OCR and Azure Document Intelligence returned the line items as a table. Google Document AI, AWS Textract and Tesseract returned them as plain text. On a scanned 25-cell table (details), the same five returned every cell correctly.
Handwriting. Word accuracy on one handwritten note (details):
| API | Word accuracy |
|---|---|
| Jina OCR | 98.9% |
| Mistral OCR | 96.8% |
| GLM-OCR | 92.6% |
| PaddleOCR-VL | 89.4% |
| AWS Textract | 85.1% |
| Azure Document Intelligence | 84.0% |
| Google Document AI | 76.6% |
| Tesseract | 6.4% |
Low resolution. With the invoice shrunk to 333 x 253 pixels (details), Tesseract read 6.9% of the words; the other seven APIs read between 89.7% and 100%.
Each of these is one file. They are good at showing which APIs fail in a category and bad at ranking the ones that pass. For public benchmark scores across many documents, see best OCR model.
Free OCR APIs
If you need OCR at no cost:
- Tesseract through an API. Tesseract is free on OCRRouter. It is fine for clean printed text and poor at everything else (6.4% on our handwriting test).
- Sign-up credits. A new OCRRouter account gets 50 credits, which is 50 pages on a $1 per 1,000 pages model such as DeepSeek-OCR or Jina OCR.
- Cloud free tiers. Google Cloud Vision, AWS Textract and Azure Document Intelligence each offer a monthly free allowance on their own platforms. You need a cloud account, and the limits and terms are on each vendor's pricing page.
- Free tools without code. For a handful of files, image to text and PDF to text run in the browser without an account.
OCR API or OCR SDK?
An SDK runs OCR on the device: Apple's Vision framework on iPhone and Mac, Google ML Kit on Android, Tesseract on a server. It works offline, has no per-page cost and keeps files on the device. It is weaker on tables, handwriting and poor scans.
An API sends the file to a hosted model. It needs a network and costs per page, and in return you get the strongest models, Markdown tables and fields without shipping a model in your app. Many apps use both: on-device OCR for live camera text, an API for documents the user saves.
Calling an API needs no special SDK. One HTTP request is enough:
import os, requests
r = requests.post(
"https://api.ocrrouter.ai/v1/ocr",
headers={"Authorization": f"Bearer {os.environ['OCRROUTER_API_KEY']}"},
json={
"model": "mistral/mistral-ocr",
"fallbacks": ["google/gemini-flash"],
"file_url": "https://example.com/invoice.pdf",
"format": "markdown",
},
)
print(r.json()["markdown"])
The fallbacks list retries a failed page on the next model, and you pay only for the model that answered. More languages are in the Python OCR guide and the API docs.
How to choose in one afternoon
- Collect 30 to 50 real pages, weighted toward your hardest ones: photos, handwriting, long tables, small print.
- Decide what correct means: exact text, a table with the right cells, or five fields from an invoice.
- Shortlist three APIs from the table above: one strong default, one cheap, and the one you use today.
- Run all three on the same pages. With OCRRouter that is the same request with a different
modelstring, or the playground in the browser. - Price it at your volume. At 100,000 pages a month, $1 and $25 per 1,000 pages is $100 against $2,500. The cheapest OCR page does the maths for every model.
- Keep a fallback so one provider's outage does not stop your pipeline.