About 500 double-spaced pages on a 200K token model, and roughly 1,900 pages on the 1M token models. PDFs are a different unit. One PDF page costs three to six times a page of pasted text, so a single API request is capped at 100 PDF pages on 200K models and 600 on 1M models.
The gap between those two numbers is where almost every answer online goes wrong. “500 pages” and “100 pages” both come from Anthropic, they are both correct, and they are measuring two different things. Once you know which one your file is, the number stops being a guess.
Anthropic says 500 pages. What size page is that?
Anthropic’s help centre states that outside its largest models, Claude’s context window is 200K, “meaning it can ingest 200K+ tokens (about 500 pages of text or more)”. It never says how big a page is, which is why the figure gets copied around and quietly changes meaning.
You can work it out from Anthropic’s own glossary, which says a token “approximately represents 3.5 English characters”. That gives 200,000 tokens x 3.5 = 700,000 characters. An average English word runs about 4.7 letters plus a space, so call it 5.7 characters a word. That is roughly 123,000 words. Divide by 500 pages and you get 246 words a page.
246 words is a double-spaced manuscript page. It is not a single-spaced A4 page, which holds closer to 500 words, and it is nothing like a dense academic page at 650. So Anthropic’s figure is honest, but it is the most generous page size in common use. Swap in your own page and the answer moves a long way.
| Kind of page | Words per page | 200K window | 500K window | 1M window |
|---|---|---|---|---|
| Double-spaced manuscript the page size Anthropic's own figure assumes |
250 | 490 | 940 | 1,890 |
| Trade paperback | 300 | 410 | 790 | 1,570 |
| Single-spaced A4 or Letter | 500 | 250 | 470 | 940 |
| Dense academic or legal | 650 | 190 | 360 | 730 |
| PDF page sent as text plus a full page image |
n/a | 67 to 133 | 167 to 333 | 333 to 667 |
| Words the window holds in total | 123,000 | 236,000 | 472,000 |
Our arithmetic from Anthropic's published figures. 200K uses the older tokenizer at 3.5 characters per token, since Claude Haiku 4.5 and earlier models sit in that tier. 500K and 1M use the Claude 4.7 and later tokenizer at about 2.7 characters per token. Words assume 5.7 characters including the space. Totals are the whole window, so leave room for your prompt and Claude's reply.
The 500K row matters because most people do not know that tier exists. Anthropic’s help centre lists three sizes for chat, not two: 1M for the newest models, 500K for the generation before them, and 200K for everything else.
Why the 150,000 words answer is now wrong
Search for this and you will be told 200K tokens is 150,000 words, on every result. That number comes from the old rule of thumb of 0.75 words a token, which OpenAI published for its own models years ago. It is no longer safe for Claude.
Anthropic’s token counting documentation says that “Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer. The same input text produces approximately 30 percent more tokens than on earlier models.” It then says plainly: when migrating from an older model, “don’t reuse token counts measured on the older model to estimate costs or context window fit”.
That 30 percent is not a rounding error. On the newer tokenizer a token carries about 2.7 characters instead of 3.5, so the same window holds less text. Run the old rule on a 1M window and you get 750,000 words. Run the real one and you get about 472,000. The popular answer overstates what a current model can actually hold by around 59 percent.
If you are sizing a prompt against a real budget, count it on the model you will actually call. Our Claude Token Counter lets you paste the text and switch models, and the tool page has the full context window and pricing table per model.
A PDF page is not a page of text
This is the part that catches people out. When you send a PDF to Claude, Anthropic’s PDF support documentation says the system “converts each page of the document into an image”, then hands Claude the extracted text and that image together. You are paying for both.
Anthropic puts the cost at 1,500 to 3,000 tokens a page, depending on how dense the page is. A 250 word page of plain pasted text costs about 530 tokens on a current model. So a PDF page costs roughly three to six times what the same page would cost as text.
| What you send | Tokens per page | Pages in 200K | What Claude can see |
|---|---|---|---|
| Text pasted into the chat | about 530 250 word page, current tokenizer |
about 375 | The words only. Layout, tables and charts are gone. |
| PDF, light page mostly prose, little styling |
1,500 | 133 | Words plus the rendered page, so layout survives. |
| PDF, dense page tables, charts, small type |
3,000 | 67 | Everything, including the graphs. This is what you are paying for. |
| Cost of a PDF page versus a text page | roughly 3x to 6x and the ratio is worse on older models, about 4x to 7x | ||
The 1,500 to 3,000 range is Anthropic's own published figure for PDF pages sent to the API. The reason it is so much higher than plain text is that the system converts every page into an image and sends that alongside the extracted text, which is also why Claude can read a chart out of a PDF but not out of pasted text.
Do the division and 200K tokens holds 67 to 133 PDF pages. 1M holds 333 to 667. Now compare that to Anthropic’s published caps: 100 PDF pages per request on a 200K model, 600 on a 1M model. The caps sit almost exactly where the token arithmetic puts them, which is a useful thing to know, because it tells you the limit is not arbitrary. It is the window.
For reference, Google states that each document page in the Gemini API is 258 tokens. A Claude PDF page therefore costs around six to twelve times a Gemini one, because Gemini is not rendering every page as a full image by default. Different trade off, not a worse one: the image is what lets Claude read a chart.
The hard limits, in the app and in the API
Context window is one ceiling. The product has separate ones, and they are the limits you will actually hit first.
| Limit | Claude app, claude.ai | API |
|---|---|---|
| PDF pages per file or request | 1,000 pages | 600 on a 1M model, 100 on a 200K model |
| Pages Claude can see as images | first 100 only. Pages 101 to 1,000 are text only | Every page in the request, up to the cap |
| File size | 500MB per file in chat, 30MB per file in Projects | 32MB per request |
| Files at once | 20 per chat. Projects are unlimited in count but must fit the window | Bounded by the 32MB request and the window |
| Images | Up to 8000 x 8000 pixels | Priced on pixel dimensions, capped per image |
| File type | Standard PDF, not password protected or encrypted | Same |
App limits from Anthropic's file upload help article, last updated July 2026. API limits from the PDF support documentation. Whichever limit you meet first is the one that stops you, and for long documents that is usually the context window rather than the page cap.
The 100 page line inside the app is the one worth remembering. Anthropic’s upload article says PDFs are limited to 1,000 pages, but that at 100 pages or fewer Claude analyses charts and images, while pages 101 to 1,000 are processed as text only. So a 400 page report will go in, and Claude will read the words, but it will not see the graphs. Nobody tells you that in the interface.
What this means for real documents
A 90,000 word novel pasted as text is about 109,000 tokens on a current model. It fits inside a 200K window with room to work. The same novel as a 320 page PDF is 480,000 to 960,000 tokens, so it needs a 1M model, and it is still over the 100 page request cap unless you are on one.
A 150 page annual report is the common failure. As a PDF it is 225,000 to 450,000 tokens, which clears 200K on its own before you have asked a single question. Either move to a 1M model, split it, or strip it to text first and lose the charts.
A 40 page contract is comfortable anywhere: 60,000 to 120,000 tokens, under the 100 page cap, and inside the visual analysis band so Claude can see signature blocks and tables properly.
Free tool
Claude Token Counter
Paste your actual document and get the real token count for the model you plan to call, not a rule of thumb. Includes image tokens, prompt caching and the API cost for that call.
Quick answers
Can Claude read a 1,000 page PDF?
In the Claude app, yes, that is the stated file limit. Only the first 100 pages get visual analysis, the rest is text only, and the extracted text still has to fit the context window. Through the API, no: a single request caps at 600 pages on a 1M model and 100 on a 200K model.
How many pages is 200K tokens?
About 490 double-spaced pages, 410 trade paperback pages, or 250 single-spaced A4 pages. As PDF pages it is only 67 to 133, because each PDF page is sent as an image as well as text.
Does Claude read the whole document or just part of it?
The whole thing, as long as it fits. There is no sampling or skimming step. If it does not fit, the request fails or older turns get dropped, which is a different problem from Claude reading selectively.
Which Claude model reads the most?
Any of the 1M token models. As of October 2026 Anthropic’s model overview lists Claude Fable 5.1, Claude Opus 5.5 and Claude Sonnet 5.5 at 1M tokens, with Claude Haiku 4.5 at 200K.
Do images and code count the same as text?
No. Images are priced on their pixel dimensions, and code tokenises less efficiently than prose because of punctuation and indentation. The tool page covers how many tokens a Claude image uses in detail.
Item #6
Sources and date
Figures checked on 4 October 2026 against Anthropic’s own documentation: the help centre article on context window size on paid plans, the docs glossary for characters per token, token counting for the tokenizer change, PDF support for per page cost and request caps, and file upload limits for the app. The Gemini comparison is from Google’s document understanding documentation. Word and page conversions are our own arithmetic from those published figures, with the assumptions stated above.
Context windows and model names change often. This page is reviewed quarterly. If a figure here disagrees with Anthropic’s documentation, their documentation is right and we have not caught up yet.
Written & verified by
B.Tech Computer Science · 8 years in web development & SEO · Bhilwara, India
About ·
LinkedIn ·
Report an error