Handwriting OCR accuracy benchmark (2026)

A one-line summary of how each tool scored on the benchmark page. The reviews go deeper on where each holds up outside this sample, on cursive, historical hands, and non-English scripts.

• Handwriting OCR — 0.9% WER (Handwriting specialist). The tool to beat. On handwriting it returned effectively the reference text, where the next-best option made roughly ten times as many errors.
• Azure Document Intelligence — 8.67% WER (Cloud document AI). A sensible default if you live in the Microsoft ecosystem and your documents are mostly printed. Not the tool for a handwriting-heavy workload.
• AWS Textract — 10.5% WER (Cloud document AI). The right default inside AWS for printed forms. The wrong tool for documents that are handwritten throughout.
• Claude (vision) — 11.2% WER (LLM vision). Good enough for prototypes and one-off transcriptions; the silent-correction risk makes it hard to ship for document volume.
• GPT (vision) — 14.4% WER (LLM vision). Convenient and competent for casual use; the same production caveats as any LLM-as-OCR approach apply.
• Google Document AI — 23.3% WER (Cloud document AI). Fine for printed text inside Google Cloud; the weakest of the big three on handwritten prose in our test.
• Transkribus — 47.7% WER untrained (Trainable specialist). Powerful for its niche, but the untrained result is unusable; budget serious training time or pick a specialist that works out of the box.
• Tesseract — 95.4% WER (Open source). A great tool for the wrong job here. On handwriting the output is noise; include it only as a baseline.

Word Error Rate is the standard OCR accuracy metric: it counts substituted, missing, and inserted words against the reference and returns a percentage, so 0% is a perfect transcription. We use one clean, legible handwritten page as a fair baseline every serious tool should handle, which means the differences reflect the engine rather than a deliberately hard sample. We also track reading-order failures separately, because whole lines out of sequence are harder to fix than isolated word errors.

The full protocol, the reference sample, and our caveats are on the methodology page. Phone apps (Pen to Print, Google Lens, Apple Live Text) are assessed separately and not WER-scored, because their input is a photo rather than a comparable scan.