Claude vs. ChatGPT for long-document work

Editor's note: This is a documentation-led edition. We have not yet run our own hands-on tests of Claude or ChatGPT. What follows rests on official documentation, published pricing and independent studies: three peer-reviewed papers whose findings we cite from their published abstracts, and one technical report read in full. All were accessed on 16 October 2026. A tested update will follow.

Anthropic and OpenAI both sell models with windows of about a million tokens through their developer APIs. In the chat apps most people use, the picture is different, and it differs between the two products.

This comparison covers what matters once a file runs to hundreds of pages: the working window, upload limits, what happens when a document overflows into retrieval, citations, export and data terms. It does not say which assistant reads better, because we have not measured that.

What each is for: the documented facts side by side

The table compares the paid personal plans most individuals would buy, plus the API.

Sources: vendor pricing pages, help centres and developer documentation, accessed 16 October 2026. US list prices before tax. Plans and limits change often.

For the wider three-way picture, including Gemini, see our Claude, ChatGPT and Gemini comparison.

Context windows: the app is not the API

The context window is the amount of text, in tokens, a model can work with in one request, including its answer. Here the two apps diverge sharply, even though their API models are close.

Claude. Anthropic's help centre lists a 1M-token chat window on paid plans for Fable 5.1, Opus 5.5, Opus 5, Sonnet 5.5, Sonnet 5 and Haiku 5.5. Fable 5, Opus 4.6 to 4.8 and Sonnet 4.6 get 500K; other models get 200K, which Anthropic puts at "about 500 pages of text or more".[4] The pricing page lists "Up to 1M, varies by model" on every plan, Free included, though Free does not include Opus.[1] Near the limit, Claude summarises earlier messages to make room, if code execution is on.[4] A summary is not the original, so re-supply key passages late in a long session.

ChatGPT. OpenAI's pricing page gives two figures per plan. Instant mode has a total window of 27K tokens on Free, 54K on Go and Plus, and 128K on Pro, which OpenAI estimates as roughly 12, 40 and 250 pages of input. Reasoning mode has 256K on Go and Plus and 400K on Pro, about 320 and 680 pages.[5] OpenAI notes that the space for your input is "smaller than the total window", because instructions, memories and reasoning share it.[5] In the API, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna each list 1.05M tokens.[7] The model family is the same; the product limit is not.

The vendors' page estimates do not match: OpenAI equates 256K tokens with about 320 pages of input, Anthropic 200K with about 500.[5][4] Count the tokens in your own document instead.

What happens to a long PDF in each app

AI Chronicle

Uploads: what actually gets read

Claude accepts PDF, DOCX, EPUB, TXT and other text formats. In chat, each file can be 500 MB, up to 20 per chat, and PDFs up to 1,000 pages. Claude analyses "both text and visual elements" in PDFs of 100 pages or fewer; from 101 to 1,000 pages it reads text only. For other formats it extracts text and cannot read embedded images.[8]

ChatGPT accepts DOCX, PDF, TXT, spreadsheets and PPTX. Documents can be 512 MB and 2 million tokens each, with a rolling cap of 80 files every three hours; Free users get three uploads a day.[9] Outside Enterprise, plans use text-based retrieval, which "extracts digital text from the file rather than its embedded images".[9] For a report whose key numbers sit only in charts, that difference matters more than the window size.

Projects and retrieval: when the document does not fit

When there is too much to hold at once, retrieval reads only the chunks a search returns, so a passage the search misses is never seen. Our RAG explainer covers why.

Claude. Projects are open to every plan, with five on Free.[20] Project files can be 30 MB each, unlimited in number, but must fit in the context window, with "text extraction only" except for multimodal PDFs.[8] On paid plans, when project knowledge nears the limit, Claude switches to retrieval and uses "a project knowledge search tool" to fetch relevant passages, expanding capacity "by up to 10x".[10] You cannot switch it off. Anthropic says accuracy "remains consistent with in-context processing".[10] That is a vendor claim we have not tested.

ChatGPT. Plus allows 25 files per project and Pro 40; project sharing is available even on personal plans.[11] OpenAI does not document, for Free, Go, Plus or Pro, how much of a long file goes into the window and how much is searched. Its one detailed account is for Enterprise. There, up to 110K tokens of uploaded text are placed directly in the window; for a single longer document, only the first 110K tokens are included and "the remainder will only be sent to the private search index".[21] The page cites a 128K-token model limit and calls the feature "under active development", so it may predate GPT-6.[21] Its warning still applies:

"…this approach can struggle with complex tasks like summarizing very large documents or comparing multiple large files."

— OpenAI Help Center, Optimizing File Uploads in ChatGPT Enterprise (2026).[21]

Pointing back to the page: citations

Claude. The API's citations feature returns the exact text behind each claim, with page numbers for PDFs, and Anthropic says citations are "guaranteed to contain valid pointers to the provided documents".[12] That covers the pointer, not the reasoning: a real quote can still be stretched. Scanned PDFs without extractable text cannot be cited, and citations cannot be combined with structured outputs.[12] In Microsoft Word, the Claude for Word add-in (Pro and above) answers questions about the open document "with clickable section citations".[13] For ordinary chat uploads, the help pages we read do not document passage-level citations.

ChatGPT. Deep research can use uploaded files and returns a report "with citations or source links", a sources-used section, and downloads in Markdown, Word or PDF.[14] In the API, the file search tool returns file_citation annotations naming the file; the retrieved passages are returned only if you request them.[15] ChatGPT for Word, available on every plan, answers questions about the open document. Its help page does not mention section citations and advises you to "check important facts, figures, citations, and edits".[22]

Export and getting your work out

Claude. Free, Pro and Max users can export conversations and account data from Settings > Privacy on the web or desktop app, not mobile. The emailed link expires after 24 hours, and exports cannot be imported into another personal account. On Team and Enterprise, members can export their own chats only if an owner allows it, and that export "doesn't include the contents of files you uploaded".[16]

ChatGPT. Signed-in Free, Go, Plus and Pro users can request an export from Data controls. Business and Enterprise users have no self-service export; requests go to the workspace owner.[17] Files saved to Library are kept separately from the chats that used them.[23]

Data terms for confidential documents

• Claude, personal plans. Chats are used for training only if you allow it, if a chat is flagged for safety review, or if you opt into a programme.[18] Anthropic's data-usage documentation gives retention of five years with the setting on and 30 days with it off.[24] A thumbs up or down stores the whole conversation for up to five years.[18]
• ChatGPT, personal plans. New chats are not used for training once "Improve the model for everyone" is off, but feedback ratings can still send that conversation for training.[17] Deleted chats are removed within 30 days, barring legal or security exceptions; project files stay until the project or file is deleted.[23]
• Business plans. Neither vendor trains on Team, Enterprise, Business or API content by default.[19][17]

Claude for Word keeps chat history in your browser and deletes back-end inputs and outputs within 30 days.[13]

What independent research says about long inputs

No independent study we found has tested GPT-6 or Claude's 5.x models on long documents. Research on earlier models points one way: fitting text in the window is not the same as using it.

• Lost in the Middle (2023). Performance "is often highest" when relevant information sits at the start or end of the input, and drops in the middle.[25]
• RULER (2024). Of 17 models claiming windows of 32K tokens or more, only half kept satisfactory performance at 32K on RULER's harder tasks, though nearly all aced simple needle retrieval.[26]
• NoLiMa (2025). When the question and the answer shared few words, so the model had to infer the link, 11 of 13 models fell below half their short-input scores at 32K tokens. GPT-4o dropped from 99.3% to 69.7%.[27]
• Context rot (2025). Chroma tested 18 models, including GPT-4.1 and Claude 4 models, and found performance degraded as input grew in every experiment. Models did markedly better on focused prompts of about 300 tokens than on full ones of about 113K.[28] Claude models more often declined to answer when unsure; GPT models more often gave confident wrong answers when similar-looking distractors were present.[28]

"Whether relevant information is present in a model's context is not all that matters; what matters more is how that information is presented."

— Kelly Hong, Anton Troynikov and Jeff Huber, Chroma, Context Rot (2025).[28]

Read these with their limits. Every model tested has since been superseded, and the tests ran mostly through APIs rather than the apps. Chroma sells retrieval infrastructure, so it has an interest in the finding, and it scored answers with a GPT-4.1 judge calibrated against human labels. Its own needle test found no notable effect of position, which differs from Lost in the Middle.[28] The safe reading is narrow: longer inputs carry more risk of a missed or wrong answer, and the size of that risk on today's models is unmeasured in public. Our AI hallucinations explainer covers why confident answers can be wrong.

Documented limits and risks

• ChatGPT Instant is small. About 40 pages of input on Plus; use reasoning mode for longer files.[5]
• Charts can be invisible: ChatGPT outside Enterprise skips embedded images; Claude skips visuals in PDFs over 100 pages.[9][8]
• Retrieval can miss passages, and in Claude you do not choose when it starts.[10]
• Outside documents can carry instructions. Anthropic warns they "can contain hidden instructions" that manipulate its Word add-in.[13]
• Business exports are limited: Claude Team exports omit uploaded files; ChatGPT Business has no self-service export.[16][17]

Who each suits, and who should skip it

These are reasoned from the documentation, not from our use.

Claude suits people who need long documents held whole in one chat on a $20 plan; readers of chart-heavy PDFs up to 100 pages; developers who need page-level citations through the API; and Word users who want answers that jump to the cited section. Read our Claude review for its research and writing features.

ChatGPT suits people whose documents fit within about 320 pages on Plus in reasoning mode; small groups sharing a project without a business plan; and anyone who wants a cited deep research report built from uploads and the web. Our ChatGPT review covers the rest of the product.

Check first, or skip, if your key evidence lives in charts, if you rely on Instant mode for long files, or if you cannot verify answers against the source yourself.

How to test it yourself

1. Set data controls first. Turn off training, or use an incognito or temporary chat.
2. Count tokens so you know whether your file fits each plan's window.
3. Plant facts at known depths. Add five harmless facts at the start, 25%, 50%, 75% and end, then ask for each and its page.
4. Ask for a chart-only figure and see whether either tool finds it.
5. Push a project past the limit, then repeat step 3.
6. Ask what changed between two long versions of a document.
7. Check every citation against the exact sentence on the page.
8. Run a data export and check what it contains.

A worked example (hypothetical)

This scenario is invented to show how the documented limits apply. No tool was run and no outputs are reported.

Priya, an analyst at a housing charity, must summarise a 420-page consultation and flag every mention of rent caps.

Fit. If it runs to about 300K tokens, it exceeds ChatGPT Plus's 256K reasoning window but fits Claude Pro's 1M window on current models.[5][4] On ChatGPT she would split it by chapter.

Charts. The key rent figures sit in charts. Because the consultation runs past 100 pages, Claude would read its text only, and ChatGPT Plus skips embedded images, so she would extract the chart pages as a separate short PDF for Claude, or transcribe the numbers.[8][9]

Check. She would ask for every rent-cap passage with page numbers, then search the PDF herself for "rent cap" to find any the assistant missed, and confirm each quoted sentence on its page.

Questions readers commonly ask

Which has the bigger context window?

In the apps, Claude: up to 1M tokens on paid plans for its newest models, against 256K in ChatGPT Plus reasoning mode and 400K on Pro, as of 16 October 2026.[4][5] In the APIs they are close, at 1M and 1.05M.[6][7]

Can ChatGPT read a 500-page PDF?

It can accept it, since documents can be up to 2 million tokens.[9] Whether it holds the whole text at once depends on plan and mode, and OpenAI does not document how personal plans handle the overflow.[5][21]

Does a bigger window mean better answers?

No. Independent studies on earlier models found accuracy drops as inputs grow, and none has tested today's models in public.[26][27][28]

Will either tool cite the page?

Claude's API citations return the quoted passage and page, and Claude for Word links to sections.[12][13] Neither chat app documents page-level citations for uploads, so ask for quotes and check them.

The next step

The window size decides what fits; how models use what fits is the harder question. Read Context windows explained: what fits and what gets lost next.

Sources and evidence note

Plans, prices, limits, features and data terms come from Anthropic's and OpenAI's pricing pages, help centres, privacy centres and developer documentation. They describe intended behaviour, not tested performance. Independent evidence comes from three peer-reviewed conference or journal papers, read at abstract level, which is where every figure used here appears, and one company technical report read in full. Prices are US list prices before tax as displayed on 16 October 2026. The worked example is invented. All pages were accessed on 16 October 2026.

Related reading

• ChatGPT vs. Gemini for spreadsheet analysis
• How to verify AI-generated research