How Much Does Claude OCR Cost? Per Page and Per 1,000 Pages

Claude OCR costs roughly $5.31 per 1,000 pages on Claude Haiku 4.5, $19.39 on Claude Sonnet 5 and $32.32 on Claude Opus 4.8, for a scanned US Letter page transcribed to text. Anthropic does not publish a per-page rate, only per-million-token prices, so every one of those figures has to be derived. The derivation matters, because the most common way people estimate it leaves out the half of the bill that costs the most.

Below is the arithmetic in full, checked against Anthropic's own documentation on July 22, 2026. If you want the complete model-by-model breakdown with the source rules, the Claude OCR pricing reference has every tier in one table.

How much does Claude cost per page for OCR?

About half a cent to three cents, depending which model you point at the page. Here is the conversion for a scanned letter-size page at 150 DPI, transcribed to text:

Sonnet 5 carries an introductory rate through August 31, 2026 that brings it down to about $12.93 per 1,000 pages, which for now makes it the best value on the full-resolution tier.

Why most Claude OCR cost estimates are too low

Because they only count the page going in.

Every current Claude model bills output at five times input. Transcribing a page is an output-heavy job: a dense page produces something like 750 tokens of text coming back. On Opus 4.8 that output is $18.75 of the $32.32 total, about 58 percent of the bill. On Haiku 4.5 it is $3.75 of $5.31, about 71 percent. An estimate built on the input price alone lands two and a half to three and a half times under the real number.

This is the same mistake that makes published Gemini OCR figures wrong, and for the same reason. We walked through that one separately in the Gemini OCR pricing reference.

The practical consequence is a lever you control. If you need six fields off an invoice rather than a full transcription, ask for the six fields. Shorter output is the single biggest saving available on any LLM-based OCR job, and it is free to implement.

How many tokens does a Claude PDF page use?

Two separate amounts, and they stack.

Anthropic converts every PDF page into an image and extracts its text, then bills both. The documentation puts the extracted text at 1,500 to 3,000 tokens per page depending on density, with image token costs applied on top of that. So a born-digital PDF is more expensive than a scan of the same document, because the scan has no text layer to bill for.

Anthropic's own Bedrock documentation shows how big the gap is. The same 3-page PDF costs roughly 1,000 tokens in text-extraction-only mode and roughly 7,000 tokens with full visual understanding turned on. That is seven times, on one file, decided by a mode flag.

How does Claude calculate image tokens?

Claude reads an image as a grid of 28 by 28 pixel patches, and each patch is one visual token. The formula is the ceiling of width divided by 28, times the ceiling of height divided by 28.

A scanned US Letter page at 150 DPI is 1,275 by 1,650 pixels, which comes out at 46 patches across and 59 down, so 2,714 visual tokens.

Worth flagging: a lot of guides still use an older rule, width times height divided by 750. That formula is retired and gives noticeably wrong answers. You can check the current one against values Anthropic publishes itself. A 200 by 200 image is 8 by 8, which is 64 tokens. A 1,000 by 1,000 image is 36 by 36, which is 1,296. A 1,092 by 1,092 image is 39 by 39, which is 1,521. All three match Anthropic's published table exactly.

Why is Claude Haiku so much cheaper for OCR?

Partly the token price, and partly because it is looking at a smaller page.

Claude has two image resolution tiers. Fable 5, Opus 4.8, Opus 4.7 and Sonnet 5 sit on the high-resolution tier, which allows a 2,576 pixel long edge and up to 4,784 visual tokens. Everything else, including Haiku 4.5 and Sonnet 4.6, caps at 1,568 pixels and 1,568 visual tokens.

A letter page that costs 2,714 visual tokens on Opus 4.8 breaks both standard-tier limits, so it gets downscaled to roughly 1,560 tokens before Haiku ever sees it. That is a 42 percent reduction in what the model is actually reading. On a clean typed memo it makes no difference. On a dense table, a 7-point footnote or a faint carbon copy, it is exactly the detail that decides whether the number comes out right.

Raising your scan DPI will not fix this either, because a standard-tier model just scales the bigger image back down again.

Is Claude cheaper than AWS Textract for OCR?

No, and not by a small margin.

AWS Textract, Azure AI Document Intelligence and Google Document AI all charge $1.50 per 1,000 pages for plain text OCR. The cheapest Claude path is about $5.31, roughly three and a half times more. Opus 4.8 at about $32 is roughly twenty-one times more. The full cross-vendor picture is in our OCR pricing per 1,000 pages table.

That gap is not an argument against Claude. It is a statement about what each tool is for. The $1.50 services were built to turn clean, repetitive, high-volume forms into characters, and they are very good at it. Claude is a reasoning model that reads a page and works out what it means, which is why it handles handwriting, unusual layouts and documents that do not match any template you prepared for. We covered where that boundary actually falls in more detail on Claude OCR.

Most teams that run both end up routing rather than choosing. Send the standard forms to the cheap meter, send the exceptions to the model, and you get the accuracy where it matters without paying twenty-one times for the pages that never needed it. If you are running that split across several providers, it is worth having something that tracks what each API is actually costing you month to month, because per-token bills have a habit of drifting well past whatever you modelled at the start.

How many pages can Claude read in one PDF?

Up to 600 pages per request on the models with a 1 million token context window, and 100 pages when the context window is under 1 million, which today means Haiku 4.5.

The 100-page figure repeated in most write-ups is that smaller-context number, not a universal Claude limit. In practice the 32 MB request cap usually bites before the page count does, which is why large files should be uploaded through the Files API and referenced by id rather than sent inline as base64.

Which Claude model should you actually use?

Sonnet 5 is the default answer for document work. It sits on the full-resolution tier, so it sees the whole page, at about $19 per 1,000 pages or roughly $13 while the introductory rate lasts.

Haiku 4.5 is the right call for clean, born-digital documents at volume, where the downscaling costs you nothing and you save four times the money. Opus 4.8 earns its roughly $32 only when the reasoning genuinely changes the answer, which in practice means messy handwriting, nested tables, or documents where you need the model to notice that something is missing.

Before you commit to any of them, run your own worst documents through and look at the output. A token price tells you what a page costs. It tells you nothing about whether the fields come out right, and on a real pipeline that is the only number that matters.