A PDF page now costs less than a screenshot of it

You have a page on your screen. A contract, a bill, a report page. You want Claude to read it.

Most people take a screenshot and paste it in. I did too. I was wrong.

Claude cuts everything you send into tokens, small pieces of text. My test page is 1,671 letters and spaces, which is 491 tokens. So one token is about three and a half letters. Pictures also become tokens.

I sent that page to Claude three ways. The screenshot cost almost ten times the text, and more than twice the PDF.

Search online and you find the opposite advice. It says to send a screenshot instead of the PDF, because a PDF page costs more. That advice was correct. On some versions of Claude it still is. But in April 2026 Anthropic changed how its newer versions read pictures, and the answer flipped.

What a token costs depends on how you pay. Developers pay per token, so the count is money. On Pro or Max you do not buy tokens. They come out of an allowance, and Anthropic has never said how much of that allowance one token uses. So I cannot tell you what these numbers do to your limit. Nobody outside Anthropic can, and anyone who says they can is guessing.

What one page costs

I took one page of writing, about 300 words. On 6 September 2026 I sent it to eight versions of Claude, three ways each.

One number moved much more than the others. The screenshot went from 1,560 to 4,706. The PDF went from 1,957 to 2,083, which is almost nothing.

If you pay per token, Opus 5 is $5.00 a million. So a thousand pages like mine is $2.46 as text, $10.42 as PDFs and $23.53 as screenshots.

Which versions are which

Newer: Opus 4.8, Fable 5, Sonnet 5, Opus 5, Fable 5.1. Older: Opus 4.6, Sonnet 4.6, Haiku 4.5. In the app, the name at the top of the chat is the only way to tell.

On 16 April 2026 Anthropic released Claude Opus 4.7. Two things changed in it: the first version to take bigger pictures, and the first to cut text into tokens a new way. Anthropic says the new way "produces approximately 30% more tokens for the same text", and that "Claude Sonnet 4.6 and earlier models use the previous tokenizer". That second change is why my text row moves. The two rows move for different reasons.

"Older" does not mean out of date. Sonnet 4.6 came out in February 2026. All three are still on sale and in use today.

Fable 5.1 counts two tokens more than the rest of its group on everything I sent it. Those two are packaging around the message.

I did not test three versions Anthropic still sells: Opus 4.7, Opus 4.5 and Sonnet 4.5. I wanted Opus 4.7 most: the change happened in it.

What to do

1. Can you select the words with your mouse? Copy and paste. On the newer versions: text 491 tokens, PDF 2,083, screenshot 4,706.
2. Do you need the layout? Attach the PDF. Do not screenshot it. It costs 2.3 times less, and Claude gets the words as words.
3. Must you send a screenshot? Make it 1,200 pixels wide first. On a Mac: Preview, Tools, Adjust Size. On Windows: Photos, Resize. In a terminal, sips -Z 1200 shot.png. Mine went from 4,706 tokens to 1,169, and I could still read every word.

Why the screenshot costs more now

Anthropic's page about images explains the price. Claude cuts a picture into 28 by 28 pixel squares, one token each. The standard group, "All other models", stops at 1,568 pixels on the longest side and 1,568 tokens. The high-resolution group, "Claude 4.7 and later models", reaches 2,576 pixels and 4,784 tokens. Those pictures "can use up to roughly three times more visual tokens". The newer versions are not wasting your money. They look more closely, which should help them read small print.

Above 2,576 pixels the price stops. So a bigger screenshot would not cost more. A smaller one costs much less, and that is the part you control.

Two things before you use my 4,706 anywhere else. All my pictures are 16:10, the shape of a Mac screen; Anthropic's formula gives 4,784 for a 16:9 picture and 4,761 for a square one. And 4,706 is not the highest price. The published limit is 4,784. My pictures stop at 4,698, because one more row of squares would be 4,785, one token too many, so the row is removed. I ran the formula on all sixteen numbers in the table above and it gets every one right, after adding the packaging each version uses: 8 tokens on Opus 4.8, Fable 5, Sonnet 5 and Opus 5, and 10 on Opus 4.6, Sonnet 4.6, Haiku 4.5 and Fable 5.1.

The file type made no difference. The same picture as a JPEG and as a PNG both counted 4,706.

Why the PDF stayed cheap

Attaching a file adds a small fixed cost. I sent one longer text to Fable 5.1 twice. Pasted in, 4,867 tokens. Attached as a file, 4,906. The difference is 39 tokens, which I subtract below. Anthropic warns that a count "may include tokens added automatically by Anthropic for system optimizations" and that "You are not billed for system-added tokens", so 39 is what the counter shows, not necessarily 39 on a bill.

Anthropic's PDF page says what happens to each page: "The system converts each page of the document into an image. The text from each page is extracted and provided alongside each page's image."

So one PDF page costs three things. Its text, a picture of that text, and the packaging. I can measure two. What is left is the picture: 2,083 - 491 - 39 = 1,553 on the newer versions, and 1,957 - 368 - 39 = 1,550 on the older ones.

I checked that against three real ten-page PDFs, on all eight versions.

The scan is the clearest test. It has no text inside it, no words you can select with a mouse, so the only thing to subtract is the packaging and everything left is picture.

A PDF page picture is about 1,545 to 1,565 tokens, on every version and every document I tried. Now compare a screenshot of the same page. On the older versions the screenshot is 1,560, very close to my 1,550, but those versions cap any picture at 1,568, so the two cannot be far apart. On the newer versions the cap goes up to 4,784, and there the PDF page is 1,553 against 4,706 for the screenshot. The Opus 4.7 change reached pasted pictures, not PDF pages.

My sum has two weak parts. I measured the 39 tokens once, on one version, with a text file and not a PDF. And I assume text inside a PDF costs the same as pasted text. Neither changes the answer, and the scan needs neither.

I do not know why PDF pages stay cheap. Anthropic never publishes the size. Its PDF page gives no number, and another page about image coordinates says pages "are rasterized to images server-side at dimensions you don't control". But that same PDF page sends you to the image table for the price, and on a Claude 4.7 version that table suggests 4,784 tokens a page. I never measured one above 1,565.

Who found this first

Most of this is not mine. Simon Willison published both halves on 20 April 2026, five months before me. He measured a PNG at 4,744 tokens on Opus 4.7 and 1,578 on Opus 4.6, and wrote: "That 3x increase for images is entirely due to Opus 4.7 being able to handle higher resolutions". In the same post a 30-page PDF came to 60,934 tokens on the newer version, 56,482 on the older, "a 1.08x multiplier, significantly lower than the multiplier I got for raw text". Since then nobody has compared the two and noticed they change places, including the 83 comments on the Hacker News thread.

Abhishek Ray measured pictures the same day and found the same jump: 1,242 to 1,561 tokens at 1,568 pixels, 3,234 to 4,766 at 2,576. He did not test PDFs. OpenRouter looked at the change a week later across a million real requests, but only text. Its data set "Excluded media (images, files, audio, video)".

Two later articles I could not reproduce. Espressio.ai, 16 June 2026 says PDF pages can be sent at "roughly 4,784 tokens per image". I sent one one-page PDF and three ten-page PDFs to all eight versions, and no page picture went above 1,565. Gaurav Bisen, 14 July 2026 says a PDF page costs more than a screenshot of it, 3,000 to 4,500 tokens against 1,500 to 2,700. That was right for the older versions, still widely used in July: there I measured 1,957 against 1,560, and his screenshot range is right too. Denser pages would explain the difference, but my densest document was 2,457.3 tokens a page, with a picture part of 1,544.7.

How I measured it

Anthropic has a free tool, count_tokens. It counts a message's tokens before you send it, without running the model. On 6 September 2026 I ran it against claude-fable-5-1, claude-fable-5, claude-opus-5, claude-opus-4-8, claude-sonnet-5, claude-opus-4-6, claude-sonnet-4-6 and claude-haiku-4-5. Each test was one message: the page as text, the PDF as a document, or the picture as an image.

I sent one text three times to see if the count changes. It came back 450, 450 and 450. That is one text on one version, not a guarantee. Anthropic calls the count an estimate.

I measured all this through the programming interface, not the apps. The claude.ai website, the desktop app and Claude Code may shrink your upload before anything is counted, and I have not tested them. There, trust the direction and not the numbers. Smaller should still be cheaper.

Now check me. If your numbers differ, send them with the version name and the date. I will publish them next to mine.

The raw log