Search for “Arabic OCR API” and most results are ranking sites that don’t say where their scores come from. This page does the narrower job a buyer needs. For each API, it shows what the vendor’s own documentation says about Arabic, what the API returns and what a page costs. Every figure was read from the vendor’s page on 15 November 2026. BroadRouter is BroadNet’s website, and BroadNet sells one of the options below (vision LLMs, not OCR), so we say where the others do better.
If you are still deciding between OCR and a vision LLM in principle, start with Arabic OCR vs vision LLMs. It explains why connected script, dots and mixed-direction lines make Arabic hard. This post assumes you know that and want to pick a vendor.
Which OCR APIs officially support Arabic?
| API | Arabic printed text | Arabic handwriting | What it returns |
|---|---|---|---|
Google Cloud Vision (DOCUMENT_TEXT_DETECTION) |
Yes, ar, listed as Modern Standard |
Not listed (Arabic script is not in the handwriting table) | Text with page, block, paragraph, word and break information |
Azure AI Document Intelligence, Read (prebuilt-read) |
Yes, ar |
Yes, in v4.0 (not in v3.1) | Printed and handwritten text |
Azure AI Document Intelligence, Layout (prebuilt-layout) |
Yes, ar |
Yes, in v4.0 | Text, tables, document structure and selection marks |
| Azure prebuilt invoice and receipt | Invoice lists Arabic (ar); thermal receipts list Arabic |
n/a | Named invoice and receipt fields |
| Amazon Textract (Detect Document Text) | Yes, among 40+ languages | No: “handwriting … in English only” | Words and lines, each with a bounding box and confidence score |
Mistral OCR (mistral-ocr-latest) |
Yes, listed under Middle Eastern languages | Not stated | Per-page Markdown plus JSON fields: tables, images, headers, footers, blocks, confidence scores |
| Vision LLM (Qwen3.8 27B, Gemma 4 VL 26B, DeepSeek V4.1 Flash on BroadNet) | Reads Arabic images; not an OCR engine | Not documented | Whatever you ask for: text, fields, JSON, a translation |
Two details matter more than the yes/no columns:
- Textract limits more than handwriting. Its FAQ says “Handwriting, Invoices, and Receipts, Identity documents, and Queries processing are in English only.” On Arabic documents, Textract gives you text and positions, not invoice fields.
- Azure’s Arabic handwriting is new in v4.0. The v3.1 handwriting table does not include Arabic. Pin the API version you test against.
Mistral’s language page says its list reflects “languages with strong expected performance”, which is a vendor statement, not a measured score.
What does each Arabic OCR API cost per 1,000 pages?
Prices in US dollars, read on 15 November 2026 from each vendor’s pricing source:
| API and feature | Free allowance | Entry price per 1,000 pages | High-volume price |
|---|---|---|---|
| Google Cloud Vision, Document Text Detection | First 1,000 units a month | USD 1.50 (units 1,001 to 5,000,000) | USD 0.60 above 5M units a month |
| Azure Document Intelligence, Read (East US) | 500 pages a month on the free tier | USD 1.50 | USD 0.60 after the first 1M pages |
| Azure Document Intelligence, Layout and prebuilt models (East US) | Same free tier | USD 10 | Commitment tiers are listed on Azure’s pricing page (prices load in the browser; check your region) |
| Amazon Textract, Detect Document Text (US West, Oregon) | 1,000 pages a month for new AWS customers, for three months | USD 1.50 (first 1M pages a month) | USD 0.60 after 1M pages |
| Mistral OCR | Not stated | Priced per 1,000 pages; the amount is not in the page’s static HTML, so we could not confirm it | Mistral says batch processing halves its API prices |
Google bills each PDF page as one image, and each feature as a separate unit. Asking for two features on one page costs two units. Azure’s prebuilt price covers layout, invoice and receipt. If you need structured invoice fields from Azure, plan on the USD 10 rate, not the USD 1.50 Read rate.
What does a vision LLM cost per page?
A vision LLM has no per-page price. On BroadNet pay-as-you-go keys, the docs say an image “counts as input tokens at the model’s input price”, and a larger image uses more tokens. The page cost is the image tokens plus your prompt, plus the answer at the output price.
Image tokens vary by model and resolution. Gemma 4 uses budgets of 70, 140, 280, 560 or 1,120 tokens per image, and Google’s docs recommend a higher budget for the best accuracy on visual tasks. As a planning shape, not a measurement, take about 1,000 input tokens and 300 output tokens per page:
| BroadNet model | Input per 1M | Output per 1M | 1,000 pages at about 1M input and 0.3M output |
|---|---|---|---|
| Gemma 4 VL 26B | $0.10 | $0.22 | $0.10 plus three tenths of $0.22 |
| Qwen3.8 27B | $0.175 | $2.17 | $0.175 plus three tenths of $2.17, plus reasoning |
| DeepSeek V4.1 Flash | $0.175 | $0.52 | $0.175 plus three tenths of $0.52, plus reasoning |
Qwen3.8 27B and DeepSeek V4.1 Flash reason before they answer, and reasoning is billed as output. Keep max_tokens at 1,000 or more, or set reasoning_effort to low, or the answer can come back empty (why). Measure your own pages: send ten, read usage.prompt_tokens and usage.completion_tokens, and multiply. The image token calculator and vision LLM cost per image do the arithmetic.
The comparison is not like for like. USD 1.50 per 1,000 pages buys raw text and positions. A vision LLM call can return the vendor, VAT number, date and total as JSON in the same request, which saves you the parsing step. Gemma 4 VL 26B is also covered by the $5 starting credit. Qwen3.8 27B and DeepSeek V4.1 Flash need a top-up.
How do the integration details differ?
| Google Cloud Vision | Azure Document Intelligence | Amazon Textract | Mistral OCR | BroadNet vision LLMs | |
|---|---|---|---|---|---|
| Input | Images; PDF and TIFF through DOCUMENT_TEXT_DETECTION |
See the Read model’s file-type list | See the Textract file-type list | image_url (PNG, JPEG, AVIF and more) or document_url (PDF, PPTX, DOCX and more) |
One image per content item as a data: URL; render PDF pages to images first |
| Language hint | Optional ar hint |
Optional; Microsoft advises not to set it unless you are sure | Not covered in the FAQ | Not stated | Say “the document is in Arabic” in the prompt |
| Structured fields | No (text hierarchy only) | Yes, prebuilt invoice and receipt | Invoices, receipts and Queries are English only | Optional document annotations | Ask for JSON; structured output is documented for Qwen3.8 27B only |
| API shape | Google Cloud API | Azure REST and SDKs | AWS SDKs | Mistral /v1/ocr |
OpenAI-compatible /v1/chat/completions |
Microsoft’s note on language codes is worth reading if your documents mix Arabic and English: “Don’t provide the language code as the parameter unless you are sure of the language … Otherwise, the service may return incomplete and incorrect text.”
Which Arabic OCR API should you choose?
There is no single answer, and no vendor publishes an independent Arabic accuracy test we could verify. Choose by output first, then test:
- You need searchable text from long scanned archives. Google Cloud Vision, Azure Read or Textract. All three list Arabic printed text at the same entry price. Choose the cloud you already use for storage and identity.
- You need handwritten Arabic. Azure Document Intelligence v4.0 is the only OCR API here that lists it. Also test a vision LLM on the same pages, since none of them document handwriting either way.
- You need invoice fields from Arabic invoices. Compare Azure’s prebuilt invoice model (USD 10 per 1,000 pages) with a vision LLM that returns JSON. On BroadNet, Qwen3.8 27B is the model with documented structured output. Read Arabic invoices with a vision model has working code.
- You want Markdown with tables for RAG. Mistral OCR returns per-page Markdown with tables and blocks. Confirm its current per-page price on Mistral’s pricing page.
- You need fields, translation and a summary in one call. A vision LLM. The pipeline, with validation, is in extract invoice data to JSON with a vision LLM.
How do you test Arabic OCR APIs on your own documents?
- Pick 50 to 100 real pages: clean PDFs, phone photos, faded thermal receipts, stamped invoices and, if you have them, handwritten forms.
- Type a ground truth for the fields you care about. Keep Eastern Arabic digits (٠١٢٣) as printed, and note them separately.
- Score characters and fields separately. Character error rate shows reading quality. Field accuracy (total, date, VAT number) shows business quality. A one-digit error in a total matters more than ten in a footer.
- Check order on mixed lines. A line such as “الإجمالي: 1,250.00 SAR” can come back with the number before the label. That is a bidirectional-order problem, not a recognition error, but your parser still has to cope with it.
- Validate numbers in code whatever you choose: line items sum to the total, VAT matches the rate and dates parse. Send failures to a human review queue.
Where BroadNet Inference fits
BroadRouter is BroadNet’s website. BroadNet Inference does not sell an OCR engine. It sells vision LLMs on one OpenAI-compatible API: Qwen3.8 27B, BroadNet Gemma 4 VL 26B and DeepSeek V4.1 Flash, all of which accept images. If you need the cheapest raw text from millions of scanned pages, or Arabic handwriting with positions, a classic OCR API listed above is a better fit. BroadNet fits when you want fields or JSON from Arabic documents, with zero data retention: BroadNet says inference content is not retained. Try Gemma 4 VL 26B on the $5 starting credit, or read the API reference and get a key.
