Best OCR for Handwriting and Complex Tables in 2026
The short answer
Handwriting and complex tables are two different problems, and the best OCR for each is different. Cloud OCR from Microsoft and Google is cheap and good at handwriting across languages, but flattens complicated tables unless you pay for a layout model. Newer vision-language OCR models such as Datalab’s Chandra 2 and Reducto’s r-1 handle both in one pass and cost more per page.
| Pick | Best for | Price per 1,000 pages (Oct 2026) | Handwriting | Complex tables |
|---|---|---|---|---|
| Chandra 2 (Datalab) | Both in one model, self-hostable | $4 Convert fast · $10 accurate · weights free under $2M revenue | Strong (notes, cursive, forms, math) | Strong (colspan/rowspan, nested headers) |
| Reducto r-1 Parse | Hardest documents, audit trail | $10 · 20% off batch | Yes | Strongest managed option |
| Azure Document Intelligence | Multilingual handwriting at scale | $1.50 Read ($0.60 above 1M) · $10 Layout | Yes, per-line handwriting flag | Layout model only |
| Google Document AI | Handwriting at volume, GCP shops | $1.50 Enterprise OCR ($0.60 above 5M) · $10 Layout Parser | Yes, handwriting hints | Layout Parser only |
| Mistral OCR 4 | Tables + bounding boxes at low cost | $4 · $2 batch | Not a headline feature — test it | Typed table blocks with boxes |
| AWS Textract | AWS-native forms and tables | $1.50 text · $15 Tables · $50 Forms (first 1M) | English only | Tables feature |
For handwriting: Azure or Google first
Azure Document Intelligence Read extracts printed and handwritten text as lines and words with bounding polygons, and marks each line as handwritten or not with a confidence score — useful for routing low-confidence handwriting to a reviewer. It costs $1.50 per 1,000 pages and $0.60 above the first million (Azure retail price list, October 2026). Handwriting support is listed per language in Microsoft’s OCR language table, so check yours.
Google Document AI Enterprise OCR costs the same $1.50 per 1,000 pages (first 1,000 free, $0.60 above 5 million) and adds language and handwriting hints plus a font-style add-on that flags handwritten words. Pick whichever cloud you already run on; on plain handwritten notes the difference between them is smaller than the difference between a clean and a poor scan.
AWS Textract reads handwriting but, per its own quota page, “handwritten character recognition is only supported in English,” and it does not read vertical text. Fine for US forms, wrong for multilingual archives.
For complex tables: Chandra 2 or Reducto
Datalab’s Chandra 2 (4B parameters, released March 2026) is the best value when tables and handwriting appear on the same page. Datalab reports 85.9% overall on the olmOCR benchmark, 89.9% on its tables category and 89.3% on old scanned math, and lists colspan, rowspan and hierarchical headers, forms with checkboxes and handwriting among its strengths. The hosted API costs $4 per 1,000 pages for Convert in fast or balanced mode and $10 for accurate mode; Datalab says the hosted model is more accurate than the open weights. The weights are licensed under a modified OpenRAIL-M: free for research, personal use and companies under $2M in funding or revenue, commercial licence above that.
Reducto r-1 Parse is the managed choice when a wrong table cell is expensive (contracts, financial statements, insurance). It reconstructs merged cells, nested headers and multi-page tables into HTML, Markdown, JSON or CSV, handles handwriting and faded scans in the same model, and returns a bounding box and confidence score per output for audit. It costs $10 per 1,000 pages, 20% less in batch.
Mistral OCR 4 (June 23, 2026) is the cheap structure-aware middle: it classifies blocks as titles, tables, equations or signatures, returns bounding boxes and per-word confidence, and costs $4 per 1,000 pages or $2 in batch. Mistral does not lead with handwriting, so test it on your handwritten pages before relying on it.
How to choose
- Mostly handwriting, few tables → Azure Read or Google Enterprise OCR at $1.50 per 1,000 pages.
- Mostly tables, typed text → Mistral OCR 4 ($2–$4) for volume, Reducto ($10) where errors cost money.
- Handwritten forms, tables with handwriting, historical records → Chandra 2 (hosted or self-hosted) or Reducto.
- Data cannot leave your network → self-host Chandra 2 (check the licence) or Azure’s disconnected containers.
- Always keep confidence scores and boxes, and send low-confidence cells to a human. No 2026 OCR is reliable enough on cursive or merged-cell tables to skip review on high-stakes fields.
Benchmark scores here are vendor-reported. Run 100 of your own worst pages through two candidates before committing.
Related: the full OCR API ranking, the best way to parse scanned PDFs with low latency, alternatives to Google Cloud Vision and OmniPage.
Last verified: October 6, 2026. Prices are list USD from vendor pricing pages and the Azure retail price API.