Skip to main content
FastOCR

OCR.space free: the web tool, the API limits and an alternative

Short answer

OCR.space is free two ways: a web tool for files up to 5 MB with no registration, and an OCR.space API whose free key allows 25,000 Engine 1/2 requests a month plus 1,000 on Engine 3. On our scans Engine 3 matched FastOCR on Arabic and trailed on Urdu and Bengali. FastOCR (we make it) is a free alternative returning text, a searchable PDF, Word and Markdown.

By the FastOCR team · Updated 4 November 2026 · Accuracy test run 31 October 2026

We make FastOCR. Accuracy figures come from our own test of each tool we scored, ours included, on the same scans; other facts link to each tool's own pages. How we tested

Try it here

Drop your file here

PDF, PNG, JPG, WebP, BMP, HEIC

PDF up to 1 GB · images up to 20 MB

Tips for a clear photo
  • Lay the page flat.
  • Use even light, with no shadows.
  • Turn off the flash to avoid glare.
  • Fill the frame with the page.

Verdict

Pick FastOCR if

  • You want finished files, not JSON: text on screen, a searchable PDF and Word (.docx) from one image upload, with no key and no watermark.
  • Your page is Urdu or Bengali, which OCR.space's language table does not name. FastOCR got 9.1% and 7.4% of characters wrong on our pages; Engine 3 got 19.6% and 11.2% (lower is better).
  • You would rather not pick an engine or a language code. FastOCR has one recognizer and detects the language, mixed pages included.
  • You also want translation, AI Polish, free in-browser PDF tools or bank statements in Excel.

Pick OCR.space if

  • You call OCR from code on many small images and want a free key that counts requests rather than pages (25,000 a month on Engines 1 and 2).
  • You want a searchable PDF with a visible text layer, or handwriting OCR and Markdown tables, which it advertises for Engine 3.

FastOCR and OCR.space side by side

Account to startFastOCRNone for images. A free account for PDFs.OCR.spaceWeb tool: "no registration required". API: a free key.Source
Free web limit, as advertisedFastOCRVisitors have a small daily image cap; a free account lifts it. See pricing.OCR.space5 MB per documentSource
Free APIFastOCRA monthly page allowance for any signed-in account, shared with web PDFs, then Page PacksOCR.space25,000 Engine 1/2 requests a month plus 1,000 Engine 3 conversions, 1 MB files, 3 PDF pages, 500 requests a day per IPSource
OutputsFastOCRText on screen, searchable PDF, Word (.docx), MarkdownOCR.spaceText, JSON from the API, and a searchable PDF with a visible or invisible text layer. Engine 3 can return tables as Markdown.Source
Choosing the languageFastOCRDetected automatically, including mixed pagesOCR.spaceAuto-detection on Engines 2 and 3Source
Languages listedFastOCR100+ languagesOCR.spaceAbout 35 named, plus over 200 more on Engine 3. Urdu and Bengali are not named; Persian, Hebrew and Tamil are listed for Engine 3 only.Source
Watermark on free outputFastOCRNone on image resultsOCR.spaceFree API searchable PDFs: "with watermark". The web tool page does not say.Source
Paid plansFastOCRMonthly plans, or one-off Page Packs. See pricing.OCR.spaceAPI plans: PRO $30 a month, PRO PDF $60 a monthSource
Your filesFastOCRProcessed on FastOCR servers; retention is in the privacy policy. The browser PDF tools never upload.OCR.spaceWeb tool deletes files and text right after OCR, on EU servers; API PDF download links last one hourSource
Share of characters wrong on the six scanned pages shown (lower is better)FastOCR1.0% (Chinese) to 9.1% (Urdu)OCR.spaceEngine 3: 2.9% (Arabic) to 19.6% (Urdu) on the 3 pages it returned. Engine 2 rejected the Arabic page.Our test

Why FastOCR

  • Photos and screenshots

    Every image upload gets raw text and a single-page searchable PDF, plus Word and Markdown.

    See Photos and screenshots

  • Scanned PDFs

    Scanned PDF in, searchable PDF out; text copies in the right order, including right-to-left.

    See Scanned PDFs

  • Languages

    100+ languages, detected automatically, including mixed pages.

    Share of characters wrong on our evidence scans (lower is better): Arabic 2.6%, Urdu 9.1%, Bengali 7.4%, Tamil 2.2%, Chinese 1.0%.

    See Languages

  • For developers

    REST API, Python SDK, MCP server for AI agents, LangChain and LlamaIndex loaders.

    See For developers

  • Offline desktop apps

    Free offline Mac and Windows apps (public beta): searchable PDF, text, Word and Markdown, nothing leaves your computer.

    See Offline desktop apps

Proof

On the pages where OCR.space Engine 3 returned text, it was level with FastOCR on Arabic, inside the 3-point noise, so that page counts as a tie; FastOCR was clearly ahead on Urdu and Bengali. The default Engine 2 trailed or matched FastOCR on every page it read (lower is better).

Three lines of a printed page of the Urdu book Tota Kahani (written 1801) in Nastaliq script, on stained paper.(opens full size)

FastOCR

9.1% of characters wrong
احسان اُس خدا کا کہ جس نے دریائے سخن کو اپنے ابر کرم سے گو ہر معنی بخشا اور زبان
کو واسطے اپنی
حمد کے گویا کیا اور پیغمبر آخرالزمان کو ہم گنہگاروں کی شفاعت کی اسے رحمتہ اللعالمین پیدا کیا کہ

OCR.space Engine 3

19.6% of characters wrong
احسان اُس خدا کا کہ جس نے دریا سے سُخن کو اپنے ابر کرم سے گوسر معنی بخشا اور زبان کو واسطہ بھیجنے کو گویا کیا اور پھر آخر مان کو ہم گنگاوں کی شفاعت کیواسطے رحمتہ اللعالمین پیدا کیا کہ

OCR.space Engine 2 (default)

30.4% of characters wrong
احسان أس مداكاكيج سنءريات سخن كواب ابركرم ← كوبر معى تختا اورزيان كوداسطيى
حدث كوياكيا اور خير آخر الزيان كويم كنبكارول كى شفاعت كيوا سطى رحمة اللعالمين بيد كياكه
The opening lines of an Urdu page (Tota Kahani, written 1801) and the first lines each engine returned for them, unedited. Each percentage is the share of characters that engine got wrong on the whole page, so lower is better. FastOCR output is from 25 October 2026; OCR.space ran on 31 October 2026 through its public demo key. Scan: Wikimedia Commons (public domain). Text: Wikisource.

% of characters wrong by script, lower is better

  • FastOCR
  • OCR.space Engine 3
  • OCR.space default Engine 2
  • imagetotext.info
Character error rate by script, lower is better
ScriptCharacter error rate by tool
Chinese
FastOCR 1.0%
OCR.space E3 no score
OCR.space E2 3.0%
Tamil
FastOCR 2.2%
OCR.space E3 no score
OCR.space E2 99.8%
Arabic
FastOCR 2.6%
OCR.space E3 2.9%
OCR.space E2 not supported
imagetotext.info 2.9%
English
FastOCR 3.3%
OCR.space E3 no score
OCR.space E2 4.5%
Bengali
FastOCR 7.4%
OCR.space E3 11.2%
OCR.space E2 100.0%
imagetotext.info 7.9%
Urdu
FastOCR 9.1%
OCR.space E3 19.6%
OCR.space E2 30.4%
imagetotext.info 9.2%
Character error rate on one public-domain book page per script. Lower is better. Gaps under about 3 points are within the noise of the reference text. "Not supported" means the tool refused the language; "no score" means it returned nothing, which says nothing about accuracy. Engine 3 returned nothing on three pages, probably because the demo key was rate limited. How we tested.

Try it on your own page

Is OCR.space free? The web tool and the API limits

Yes, OCR.space is free in two ways. The web tool at ocr.space (often typed as ocr space or ocrspace) takes an image or PDF of up to 5 MB, needs no registration, and returns text or a searchable PDF with a visible or invisible text layer. The free API is a separate plan: you register for a key and get 25,000 requests a month on Engines 1 and 2 plus 1,000 Engine 3 conversions, files up to 1 MB, 3 pages per PDF and 500 requests a day from one IP address. Searchable PDFs made with the free key are marked "with watermark" in its plan table. Paid API plans start at $30 a month for PRO and $60 for PRO PDF, and the pricing lists no web subscription.

The engine you run matters more than the limits. OCR.space names three: Engine 1 (deprecated), Engine 2 (the default) and Engine 3, its newest, which it says handles handwriting and over 200 extra languages. Its language table names about 35 languages. Urdu and Bengali are not among them, and Persian, Hebrew and Tamil appear for Engine 3 only. Both newer engines offer automatic language detection. Engine 3 can return tables in Markdown format. For the web tool, OCR.space says files and extracted text are deleted right after OCR and that its servers are in the EU.

A free OCR.space alternative: what FastOCR adds

FastOCR is a free OCR.space alternative for people who want usable files rather than JSON. Drop an image into the free OCR online tool and you get the text next to the page, a single-page searchable PDF, Word (.docx) and Markdown, with no sign-up and no watermark. A scanned PDF comes back searchable with a free account, and its text copies in reading order, right-to-left Arabic and Urdu included. There is no engine menu and no language code: it detects 100+ languages on its own, including pages that mix scripts, and Urdu and Bengali are covered.

When the text is out, you can translate it or tidy it with AI Polish; both work on finished images. The site also has free PDF tools that run in your browser, from merge and split to compress and sign, and a bank statement converter that writes Excel, CSV and QuickBooks files. For OCR without a cloud API, the free Mac and Windows apps (public beta) work offline; they use the text recognition built into each system, so the accuracy numbers here do not apply to them.

Photo of a page

Photo of an old printed English book page, slightly grey and uneven, starting with the words 'ing, and Haggart gained an opportunity'.(opens full size)

Text FastOCR returned

3.3% of characters wrong
ing, and Haggart gained an opportunity
to escape; he left Perth for Glasgow, and
made up his mind to go to Ireland, and
started in the Steam Boat; here he wa
recognised by the Provost of Kircaldy
who went ashore at Lamlash; Haggar
The first lines of a page FastOCR read on 25 October 2026, shown as it returned them; words cut at the page edge stay cut. The percentage is the character error rate: the share of characters wrong, missing or extra, so lower is better. Scan: Wikimedia Commons (public domain). Text: Wikisource.

OCR.space vs FastOCR: accuracy on ${shownCountWord()} scripts

OCR.space Engine 3 was level with FastOCR on Arabic: 2.9% of characters wrong for Engine 3 and 2.6% for FastOCR, a gap inside our 3-point noise band, so we treat it as a tie. The score is character error rate: the share of characters an engine got wrong, dropped or added, so lower is better.

FastOCR led Engine 3 clearly on Urdu and Bengali: Urdu 9.1% against 19.6%, and Bengali 7.4% against 11.2%. Engine 3 returned nothing for Tamil, Chinese and English. That may have been the public demo key's rate limit, so we show no result and do not count it against OCR.space. Lower is better throughout.

The default Engine 2 did worse. It rejected the Arabic page outright, got 30.4% of the Urdu page wrong, and almost every character on Bengali (100.0%) and Tamil (99.8%). On Chinese and English it stayed inside the noise: 3.0% and 4.5% against FastOCR's 1.0% and 3.3% (lower is better). OCR.space's language table lists Arabic for Engine 2, so a run today with a key of your own may go better than our October demo-key run did.

OCR.space API vs the FastOCR API

Both are REST APIs with a free tier, but they meter different things. OCR.space counts requests: 25,000 a month on Engines 1 and 2 plus 1,000 on Engine 3, each file up to 1 MB and each PDF up to 3 pages, with the text returned as JSON and a searchable PDF on request. Its API page documents no parameter for choosing which PDF pages to read, so split the file locally and send only the pages you need; FastOCR's Split PDF tool does that in the browser.

FastOCR counts pages. Any signed-in account can create a key, gets a free page allowance each month (the same one web PDFs use) and can buy Page Packs after that. Send a PDF or an image, then download plain text or a searchable PDF from the same job. There is a Python SDK, an MCP server for AI agents, and LangChain and LlamaIndex loaders. The API docs walk through the calls.

How we tested OCR.space

Test pages: public-domain book scans from Wikimedia Commons, one each in Arabic, Urdu, Bengali, Tamil, Chinese and English. We compared every output with the Wikisource transcription of the same page, after removing spaces, punctuation and Arabic diacritics. Those transcriptions differ from a hand-checked key by 1% to 6%, which is why a gap under about 3 points counts as a tie. The benchmark write-up has the wider picture.

OCR.space run: 31 October 2026, once per engine, through its public demo API key, on 1,280-pixel-wide scans. FastOCR run: the live API on 25 October 2026, from 500-pixel copies of the Arabic, Urdu and English pages. Blank Engine 3 results are shown as no result, not as failures. Plans, limits and languages come from ocr.space and ocr.space/ocrapi, checked on 4 November 2026. Scans: Wikimedia Commons. Text: Wikisource.

Choose OCR.space if

  • You send thousands of small images from code each month and a request quota suits you better than a page allowance.
  • You want to choose between a visible and an invisible text layer in the PDF.
  • You need handwriting OCR, which OCR.space advertises for Engine 3.
  • You want files deleted right after OCR, on servers in the EU.

FAQ

Is OCR.space free?

Yes. The web tool needs no registration and takes files up to 5 MB. A free API key allows 25,000 requests a month on Engines 1 and 2 and 1,000 on Engine 3, with 1 MB files, 3 PDF pages and 500 requests a day per IP address. Paid API plans start at $30 a month.

How many PDF pages can the OCR.space free API read?

Three. Its plan table lists a 3-page PDF limit and a 1 MB file limit on the free key. Longer PDFs need a paid plan, or splitting into parts first; its API page documents no page-selection parameter. FastOCR's free in-browser Split PDF tool can do the splitting without uploading the file.

Which OCR.space engine is best for Arabic?

Engine 3, on our test. The default Engine 2 rejected our Arabic page, while Engine 3 read it and got 2.9% of characters wrong against FastOCR's 2.6%, a tie inside the 3-point noise (lower is better). Choose the engine by hand.

Do OCR.space PDFs have a watermark?

Searchable PDFs from the free API do. Its plan table lists searchable PDF creation on the free key as "with watermark". The web tool page does not mention one either way. FastOCR adds no watermark to image results.

Does OCR.space detect the language automatically?

Yes. Its language table lists automatic detection for Engines 2 and 3. It still names only about 35 languages, without Urdu or Bengali. FastOCR also detects the language itself, including pages that mix scripts.

Is there a free OCR API besides OCR.space?

Yes. FastOCR has one: any signed-in account can create a key and gets free pages every month, then buys Page Packs. It takes PDFs and images and returns text or a searchable PDF, with a Python SDK, an MCP server and LangChain and LlamaIndex loaders.

Can I run OCR without a cloud API?

Yes. FastOCR has free offline Mac and Windows apps in public beta. They read text with the recognition built into macOS and Windows, so nothing is uploaded, and save searchable PDFs, text, Word and Markdown. Our accuracy figures cover the website, not the apps.

Method

OCR.space facts come from ocr.space and its API page, checked on 4 November 2026; anything those pages do not state is shown as Not stated, never as No. We measured accuracy on 31 October 2026 as the share of characters each engine got wrong on one public-domain scan per script, compared with the Wikisource text; lower is better. Only the tools shown with a percentage were run; other facts come from vendor pages. Details are in how we tested.

Change log

  • 12 July 2026: First published.
  • 31 October 2026: Added the script-by-script accuracy test of OCR.space Engines 2 and 3.
  • 1 November 2026: Re-checked the OCR.space plans and limits.
  • 4 November 2026: Moved to the comparison template; re-read the OCR.space home and API pages and added the API comparison.