Handwriting OCR accuracy benchmark · Updated 2026

The best handwriting OCR in 2026: an accuracy benchmark

Every "best OCR" list has an opinion. We have a number. We put the same handwritten page through 11 tools, from purpose-built specialists to cloud APIs, LLM vision, and the free phone apps, then scored each one against a hand-checked reference. Word Error Rates ran from 0.9% WER to over 95%.

Top pick

Handwriting OCR — 0.9% WER

The tool to beat. On handwriting it returned effectively the reference text, where the next-best option made roughly ten times as many errors.

Try Handwriting OCR →

The ranking

All measured tools, ranked by accuracy

Every tool below was run through the same benchmark: one handwritten English prose page, scored by Word Error Rate against a manual reference. Lower is better.

#	Tool	Type	Accuracy (WER)	Score	Price	Best for
1	Handwriting OCR Purpose-built specialist. The most accurate tool in our test.	Handwriting specialist	0.9% WER	9.6	Free 5-page trial, then from $0.15/page	Anyone with real handwritten volume: archives, cursive, historical documents, or an app that needs an accurate handwriting API.	Visit →
2	Azure Document Intelligence The strongest of the big-three cloud document APIs on handwriting.	Cloud document AI	8.67% WER	6.5	From ~$1.50 per 1,000 pages (Read tier)	Teams already on Azure processing mostly printed documents, with handwriting as a secondary case.	Visit →
3	AWS Textract Handwriting on forms, native to the AWS stack.	Cloud document AI	10.5% WER	6.2	From ~$1.50 per 1,000 pages (text detection)	AWS-native teams processing printed forms that occasionally contain handwritten fields.	Visit →
4	Claude (vision) A frontier LLM that reads handwriting well enough for prototypes.	LLM vision	11.2% WER	6.0	Token-based, roughly $5-$15 per 1,000 pages	Developers prototyping a feature who want to reuse an LLM API they already have.	Visit →
5	GPT (vision) Capable general-purpose vision, same trade-offs as any LLM.	LLM vision	14.4% WER	5.6	Token-based, roughly $2-$10 per 1,000 pages	Quick, checkable transcriptions inside an app that already calls the OpenAI API.	Visit →
6	Google Document AI Capable on print; a reading-order problem on handwritten prose.	Cloud document AI	23.3% WER	4.5	From ~$1.50 per 1,000 pages (OCR)	GCP-based teams working with printed documents rather than handwriting.	Visit →
7	Transkribus Built for trained historical-document projects, not one-off use.	Trainable specialist	47.7% WER untrained	4.2	Free plan, paid from ~€19.99/mo	Research institutions with a large archive in a single consistent hand, willing to invest in training.	Visit →
8	Tesseract Excellent open-source OCR for print. Not viable for handwriting.	Open source	95.4% WER	1.5	Free and open source	Developers doing printed-text OCR who want a free, self-hosted engine. Not handwriting.	Visit →

Word Error Rate on a single legible handwritten English prose sample (2026). The phone apps (Google Lens, Apple Live Text, Pen to Print) are reviewed separately as their natural input is a phone photo, not a comparable scan. Full protocol on the methodology page.

Phone apps

The apps, tested separately

We review the phone tools apart from the benchmark table above, because their natural input is a photo rather than a comparable scan. They are convenient for tidy notes and weaker on cursive and volume.

#	Tool	Type	Score	Price	Best for
1	Pen to Print The best-known standalone handwriting app.	Mobile app	8.0	Free tier, subscription from ~$4.99/mo	Someone converting the occasional handwritten note on their phone who wants a dedicated app rather than a website or API.	Visit →
2	Google Lens Free and convenient for tidy notes; not for cursive or volume.	Free phone tool	4.0	Free	A one-off snapshot of a tidy handwritten note when you have a phone and no budget.	Visit →
3	Apple Live Text Built into iOS; convenient, limited, print-oriented.	Free phone tool	3.8	Free	iPhone and iPad users grabbing a short piece of clear handwriting on the fly.	Visit →

How to choose

It comes down to one question: printed or handwritten?

If your documents are printed, the big cloud APIs (Azure, AWS Textract, Google) are accurate and cost about $1.50 per 1,000 pages. Nothing on this page beats them on price for typed text, and you do not need a specialist.

If your documents are handwritten, those same APIs land at 8% to 23% Word Error Rate, and a purpose-built engine like Handwriting OCR brings that under 1%. On handwriting the accuracy gap is the whole story, and it pays for itself on the first page you would otherwise retype.

For a quick one-off on your phone, the free tools are fine on tidy writing. For cursive, historical hands, or any real volume, they are not.

FAQ

Common questions

What is the most accurate handwriting OCR?

In our testing, Handwriting OCR was the most accurate, at 0.9% WER on the same handwritten sample where the next-best tool made roughly ten times as many errors. The full ranked table and methodology are above.

How do you measure accuracy?

We run the same handwritten page through every tool and score the output against a hand-checked reference using Word Error Rate (WER), the standard OCR metric. Lower is better. Full protocol on our methodology page.

Is there a good free handwriting OCR?

The free phone tools (Google Lens, Apple Live Text) handle tidy, print-style handwriting but drop off on cursive and cannot process volume. Most specialist tools offer a free trial rather than a free tier. See best free handwriting OCR.

Can ChatGPT or Claude read handwriting?

Yes, both are usable: Claude vision scored 11.2% WER and GPT vision 14.4% on our sample. They are fine for prototypes and checkable one-offs, but they silently "correct" words they are unsure about, which is risky for unattended document volume.

Which is best for old letters and historical documents?

Specialist engines with explicit historical coverage lead here; general document OCR and LLM vision both fall below 50% accuracy on old scripts. See best handwriting OCR for genealogy.