Can ChatGPT Do OCR? How It Reads Images and Scanned PDFs, and Its Limits

Yes. ChatGPT reads text in images: attach a photo or a screenshot, ask for the text, and a vision model types out what it sees. OpenAI's API guide says vision-capable models can read visible text in images. The same guide lists the limits. It says vision models can make mistakes, may not perform well on non-Latin alphabets such as Japanese or Korean, and may misread rotated or upside-down text. It suggests enlarging small text. ChatGPT has no separate OCR mode to switch on, and OpenAI's pages do not name an OCR engine behind it. Source: OpenAI vision guide, checked 4 November 2026.

What that means in practice:

• Good fit: a clear photo, a screenshot or a short page, when you want the gist or a quick copy you will read through anyway.
• Check every line: the text comes back as a written reply. A model can tidy a phrase, drop a line or reorder a table while answering, and there is no confidence score or box on the page to check it against. Names, numbers, quotes and citations need a proofread.
• What it does not return: a searchable copy of your scan, or a Word or Markdown file of the page.

The steps that give ChatGPT the best chance of a faithful copy: 1. Crop the photo to the text and keep it upright. OpenAI notes that rotated text can be misread. 2. Attach the image or PDF with the + button in the message box. 3. Ask for a transcription, not a summary: "Copy all the text in this image exactly as written. Do not correct, translate or leave anything out." 4. For a PDF, ask for one page at a time, so a long reply is not cut short. 5. Read the reply against the page, line by line, before you use it.

If you would rather skip the checking, run the page through an OCR tool first and paste the result into ChatGPT. Our OCR for ChatGPT page shows that workflow with plain text and Markdown.

Often, if the PDF is short. OpenAI's help page says ChatGPT can analyze both text and images in PDFs, including charts and diagrams, and that you then ask questions about the contents. That is a promise about answering questions, not about a full transcript of a long scan. Source: OpenAI help center, checked 4 November 2026.

Why ChatGPT can't extract text from a PDF, using the limits OpenAI publishes:

• Free users are limited to 3 file uploads per day. Everyone is limited to 80 files every 3 hours, and OpenAI may lower these limits at peak hours.
• Text and document files are capped at 2 million tokens per file, and documents at 512 MB.
• PDFs saved as project files or GPT knowledge are handled differently from a PDF attached to a chat, and embedded images may not be included automatically.
• A long scanned book does not fit in one answer, so you end up asking for it page by page and checking each reply.

If you see "No text could be extracted from this file", the PDF may be a scan with no text layer under the picture, though OpenAI does not document the cause. Send the page as an image instead, or OCR the PDF first. PDF to text returns the text of every page along with a searchable PDF.

An OCR engine returns the text it recognised on the page. It is not answering a question, so it has no reason to rephrase. That matters in six situations:

• Exact, unchanged text. Contracts, invoices, quotes and references need the words as printed, not a fluent version of them.
• A searchable PDF. FastOCR gives every image a searchable PDF and turns scanned PDFs into searchable PDFs, so you can search, select and copy inside the original page.
• Long PDFs. One upload returns the whole document, page by page, as text and a searchable PDF, not as a chat reply you request page by page. PDFs need a free account, which includes a monthly page allowance; pricing lists plans and page packs for more.
• Right-to-left scripts. FastOCR reads 100+ languages and detects them itself. In its searchable PDFs, Arabic, Urdu, Persian and Hebrew copy out in the right order. OpenAI's guide names non-Latin alphabets as a weak spot and says nothing specific about right-to-left text.
• Privacy. Both ChatGPT and any OCR website process your file on their servers. When a document must stay on your computer, the free FastOCR apps for Mac and Windows, in public beta, work offline. They use the text recognition built into macOS and Windows, so the test numbers on this page are for the website, not the apps.
• Agents and code. The FastOCR API returns text and a searchable PDF, with a Python SDK, LangChain and LlamaIndex loaders, and an MCP server that lets an AI agent call real OCR on a local PDF or image.

We have not measured ChatGPT, so this page gives it no error rate. OpenAI's own guide is the best published answer: vision models can make mistakes, and non-Latin alphabets, small text and rotated text are named as weak spots. If accuracy matters, test it on one of your own pages and compare the reply with the original, character by character.

What we did measure is dedicated OCR. On 31 October 2026 we scored free OCR tools on public-domain scanned book pages, one per script, against each page's Wikisource text; FastOCR's output came from its production API on 25 October. This page shows six scripts: Arabic, Urdu, Bengali, Tamil, Chinese and English. Even here the spread is wide. With the right language pack, Tesseract got 55.3% of the Urdu page's characters wrong and 20.4% of the Arabic page. FastOCR got 9.1% and 2.6% wrong. Lower is better.

All the numbers are in the free OCR tools comparison and Arabic OCR tools.

The split that works: let OCR get the words, then let ChatGPT do what it is good at, such as summarising, rewriting, answering questions or translating. Read the OCR text against the page once, then paste it into the chat. FastOCR returns the text, a searchable PDF, Word and Markdown from one upload, so you check the words once and give ChatGPT a clean copy. ChatGPT then works from text you have already checked instead of guessing from pixels, and a clean paste is shorter than a stack of page images.

Two pages cover this in detail: OCR for ChatGPT for pasting into a chat, and PDF OCR for LLMs for pipelines that feed scanned PDFs to a model. To stay in one place, FastOCR has AI Polish to clean up recognised text and Translate; both work on a finished single image, not on PDFs.

OpenAI's API has no separate OCR endpoint. You send an image to a vision-capable model and ask it to read the text, so the result is a model reply with the same caveats as in the chat app. Source: OpenAI vision guide.

An OCR API returns recognised text instead. FastOCR's REST API returns the text and a searchable PDF, and its free monthly pages are shared with PDFs on the website. The Python SDK wraps it, LangChain and LlamaIndex loaders turn scans into documents for retrieval, and the MCP server gives AI assistants and coding agents that support MCP a tool for real OCR. The agent hands over the file and gets back the text on the page.