Tesseract OCR Software: Python Accuracy, Engine Limits and What It Really Costs

Is Tesseract a Google product?

No, and it has not been developed by Google since August 2017. This matters more than it sounds, because a large share of the people searching for this engine search for it as "Google Tesseract", and buy on the assumption that a hyperscaler stands behind it. Tesseract's own README settles it in one sentence, verbatim: "In 2005 Tesseract was open sourced by HP. From 2006 until August 2017 it was developed by Google." Before that it was Hewlett-Packard's, built in Bristol and in Greeley, Colorado between 1985 and 1994.

So the accurate description is an HP research engine that Google sponsored for eleven years and then handed back to the community. There is no Google support contract behind it, no Google SLA, and no relationship to Google Cloud Document AI, which is an entirely separate paid product. If a vendor or a consultant tells you that you are getting "Google's OCR" for free, that stopped being true nine years ago.

Is Tesseract OCR still maintained in 2026?

Yes, and the repository numbers are healthy for a project of its age. Here is the state of it, pulled from the GitHub API on September 2, 2026.

What output formats does Tesseract produce?

Seven, and the README names them exactly: plain text, hOCR (HTML), PDF, invisible-text-only PDF, TSV, ALTO and PAGE. For a developer the TSV mode is the one that matters, because it returns per-word bounding boxes and a confidence value, which is the raw material for everything you might build on top. What is missing from the list is the important part.

Can Tesseract OCR extract tables?

Not as structured data, and there is a clean way to demonstrate that rather than assert it. We parsed all ten tables in the OmniDocBench README on September 2, 2026 and counted where Tesseract appears. It is in the 12-entry text recognition evaluation and in the version list, as 5.5. It is absent from the 33-entry overall document-parsing leaderboard and absent from the 8-entry table recognition evaluation, where PaddleOCR, RapidTable, StructEqTable, GOT-OCR, Qwen2-VL and InternVL2 all appear.

That absence is not an oversight by the benchmark authors. It follows from the output list above. A table score measures how correctly a system reproduces a cell grid, and Tesseract does not emit one, so there is nothing to score. You get text and coordinates, and turning coordinates back into rows and columns is your code, your merged-cell edge cases and your silent data corruption when a column shifts. If table structure is the thing you actually need, our PDF table extraction page covers what the alternatives return, and PDF to CSV API covers the reconstruction gotchas in detail.

The limits Tesseract documents about itself

The most credible source on what an engine cannot do is usually the engine's own documentation. These are quoted verbatim from Tesseract's FAQ and its image quality guide.

Read those together and a pattern appears: Tesseract expects a clean, upright, well-cropped, 300 DPI scan of printed text. That is a reasonable expectation for a digitization project where you control the scanner. It is not a description of the documents that arrive from vendors, customers and phone cameras. If handwriting is part of your workload, handwriting OCR covers which engines actually support it.

Tesseract OCR in Python, and what pytesseract actually is

pytesseract is a thin wrapper that shells out to the Tesseract binary and hands you back a string. It is not an implementation of Tesseract and it does not bundle one, which is the single most common source of confusion. You install the engine through your operating system's package manager first, and only then does the Python package have anything to call. If you have ever seen TesseractNotFoundError, that is what it means.

The same shape applies in other languages: there are community wrappers for C#, Java, PHP and JavaScript, all of them calling the same C++ binary and all of them subject to the same output ceiling described above. Our OCR API for Python page covers what the managed alternative looks like in the same language, and OCR API for C# does the same for .NET teams.

Tesseract OCR questions, answered

Is Tesseract a Google product?

No, and it has not been developed by Google since August 2017. Tesseract's own README states that it was built at Hewlett-Packard Laboratories in Bristol and Greeley, Colorado between 1985 and 1994, open sourced by HP in 2005, and developed by Google only from 2006 until August 2017. It is a community project today. There is no Google support contract and no Google SLA behind it.

How accurate is Tesseract OCR?

On clean English text it is genuinely competitive. The independent OmniDocBench text evaluation scores Tesseract 5.5 at 0.096 edit distance on English, against 0.071 for PaddleOCR and 0.070 for OpenOCR, and it beats EasyOCR at 0.26. Accuracy falls apart elsewhere: Chinese scores 0.551 and a page rotated 270 degrees scores 0.969, which is close to total failure.

Is Tesseract OCR good enough for invoices?

For reading the words, usually yes. For extracting invoice fields, no, because it does not produce them. Tesseract outputs plain text, hOCR, PDF, TSV, ALTO and PAGE, and none of those gives you an invoice number, a vendor name or a line-item table as named values. You would be building the field extraction layer yourself on top of raw text and word boxes.

Can Tesseract OCR extract tables?

Not as structured data. Tesseract is absent from OmniDocBench's table recognition evaluation entirely, which scores eight systems, and the reason is in its own documentation: none of its output formats carries a populated cell grid. You get text and word bounding boxes, so reconstructing rows and columns is your code to write and your merged cells to get wrong.

Does Tesseract OCR use a GPU?

No. The words GPU, CUDA and OpenCL appear zero times across Tesseract's core documentation, and the FAQ answers the speed question purely in terms of CPU threads: "Tesseract 4 also uses up to four CPU threads while processing a page." That is good news for your bill. A 4 vCPU c7i.xlarge costs $0.1785 an hour, roughly a quarter of the GPU box an open-weight vision model needs.

How much does it cost to run Tesseract OCR?

The licence is free and the hardware is the cost. On AWS in US East, a 4 vCPU c7i.xlarge is $0.1785 an hour, or $130.31 a month always on. Against a managed text OCR API at $1.50 per 1,000 pages, self-hosting breaks even at about 119 pages an hour sustained. Below that you are paying for an idle instance, and none of this counts engineering time.

What is pytesseract and do I need it?

pytesseract is a thin Python wrapper that shells out to the Tesseract binary and returns the result as a string or a dataframe. You still have to install Tesseract itself first, which is the step that trips most people up. It does not add accuracy or capability. It saves you writing subprocess calls and parsing the TSV output by hand.

Why are my Tesseract OCR results so bad?

Almost always image quality rather than a bug. Tesseract's own guidance says it "works best on images which have a DPI of at least 300 dpi", that noise surviving binarisation "can cause accuracy rates to drop", and that both a missing border and an oversized border cause failures. Add rotation to that list: a page fed in at 270 degrees scores 0.969 edit distance, which is essentially unusable output.

Can Tesseract read handwriting?

Barely, and the project says so itself. The Tesseract FAQ answers this verbatim: "You can, but it won't work very well, as Tesseract is designed for printed text. Look for projects focused on handwriting recognition." If you are processing forms with handwritten entries, this is the wrong engine and no amount of preprocessing fixes it.

Can Tesseract read barcodes?

No. The FAQ answer is two sentences long: "No. Tesseract is for text recognition." If your documents carry barcodes or QR codes you need a separate barcode library alongside whatever OCR you choose.

What output formats does Tesseract produce?

Seven, per its README: plain text, hOCR (HTML), PDF, invisible-text-only PDF, TSV, ALTO and PAGE. The TSV output is the useful one for developers because it carries per-word bounding boxes and a confidence value. What is not on that list is JSON, and more importantly, any format in which Tesseract hands you named business fields.

Is Tesseract OCR still maintained in 2026?

Yes, actively. The repository has 76,309 stars and 10,785 forks, version 5.5.3 shipped on July 24, 2026, there have been 36 releases, and code was pushed on September 2, 2026. Google leaving in 2017 did not end the project. It has 485 open issues, which for a codebase of this age and reach is unremarkable.

Is Tesseract free for commercial use?

Yes, under Apache License 2.0, with no revenue threshold and no commercial tier. That is cleaner than several newer open-weight readers: Marker's model weights are free only below $5M revenue, and MinerU requires a commercial licence above USD 20 million in monthly revenue. On licence terms alone Tesseract is the safest open option available.

How many languages does Tesseract support?

The README says it can recognize "more than 100 languages" out of the box, each as a separate traineddata file you install. Treat the count as coverage rather than quality. The measured English score of 0.096 and Chinese score of 0.551 on the same benchmark show how far per-language performance can diverge, so test the scripts you actually receive.

Tesseract vs PaddleOCR: which is better?

PaddleOCR, on accuracy, and it is not close outside clean English. English is 0.071 against 0.096, which is a small gap, but Chinese is 0.055 against 0.551 and a 270-degree rotated page is 0.285 against 0.969. PaddleOCR also has a document parsing model that lands at the top of the overall leaderboard. Tesseract wins on install simplicity, on CPU-only operation and on licence clarity.

Should I use Tesseract or a managed OCR API?

Use Tesseract when you need text off clean, upright, printed English pages at high volume and you have engineers to own it. Use a managed extraction API when you need named fields, table structure, confidence per value and someone accountable when a number is wrong. The break-even is not really the hourly rate, it is who maintains the pipeline.