Anthropic sells no OCR product and publishes no per-page price. The Claude API bills tokens, and a PDF page is billed twice: once for the extracted text layer and again for the rendered page image. On a US Letter invoice that works out to about $15.83 per 1,000 pages on Claude Sonnet 5, about $6.77 on Claude Haiku 4.5, and about $39.57 on Claude Opus 5. Below is every current rate and the full derivation so you can rerun it on your own scans.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
Every other vendor in this market sells a document endpoint with a rate card. Anthropic does not. There is no OCR SKU, no prebuilt invoice model, and no page-based tier, so the only way to know what reading an invoice costs is to count tokens yourself. Four things make that harder than it looks.
AWS has Analyze Expense at $10 per 1,000 pages. Azure has a prebuilt invoice model at the same rate. Mistral publishes $4 per 1,000. Anthropic publishes a price per million tokens and nothing else, so every per-page figure on this page is derived rather than quoted.
Anthropic states that a PDF page costs text tokens (typically 1,500 to 3,000 depending on density) and image tokens, because each page is also converted to an image. Both land in the same input bill. Most cost estimates people publish count only one of the two and land roughly half too low.
Claude reads images in 28 by 28 pixel patches. Scan the same invoice at 300 DPI instead of 150 and the image half of the bill roughly quadruples, with no gain in field accuracy on a clean document. Nothing in the API warns you about this.
Output tokens are billed at five times input on every current Claude model. A verbose extraction schema costs more than the scan you sent. On Claude Haiku 4.5 the JSON you get back is over half the total cost of the page.
We took Anthropic official published token rates and its official image tokenization formula, applied them to a standard US Letter invoice scanned at 150 DPI, and added a realistic structured extraction request. Every input is shown so you can substitute your own numbers.
Anthropic documents image cost as ceil(width / 28) x ceil(height / 28) visual tokens. A US Letter page at 150 DPI is 1,275 by 1,650 pixels, which gives 46 x 59, or 2,714 visual tokens per page.
The same formula reproduces all four worked examples Anthropic publishes: 200x200 gives 64 tokens, 1000x1000 gives 1,296, 1092x1092 gives 1,521, and a downscaled 4K image gives 4,784. The math below is not an approximation.
Claude 4.7 and later models read up to 2,576 pixels on the long edge and 4,784 visual tokens. Older models cap at 1,568 pixels and 1,568 visual tokens. A full page hits the cap on the older tier, which is why Claude Haiku 4.5 is cheaper per page than the ratio of its token price suggests.
The Message Batches API applies a flat 50 percent discount to all token usage. If your invoices do not need a synchronous answer, every figure on this page halves.
Because a PDF bills text plus image and a PNG bills image only, rasterizing a page and sending it as an image is cheaper. On Claude Sonnet 5 that is $12.83 per 1,000 pages instead of $15.83, and you give up the embedded text layer.
Header fields are the easy part of an invoice. The cost that matters is per correctly extracted document, not per page, and a model that reads the header perfectly but drops half the line items is not cheaper.
Four steps to a defensible number for your document mix, using only figures Anthropic publishes.
Take the actual pixel width and height of one representative page, not the paper size. Apply ceil(width / 28) x ceil(height / 28) to get visual tokens.
Tip: If the long edge exceeds your model tier limit, Claude downscales before charging, so the token count caps out rather than growing without limit.
Anthropic gives 1,500 to 3,000 tokens per page for the extracted text, depending on density. Invoices with long line item tables sit at the top of that range.
Tip: Send a rasterized image instead of the PDF and this line disappears entirely, at the cost of losing the embedded text.
Add your system prompt and extraction instructions to input. Then estimate the JSON you expect back and add it to output, which is billed at five times the input rate.
Tip: Use the token counting endpoint on a real document rather than estimating. It is free and it settles the argument.
Multiply input tokens by the model input price and output tokens by the output price, both per million. Then compare the total against a flat monthly plan at your real volume.
Tip: Remember to add your own engineering time for retries, validation, and the line item reconciliation the raw API does not do for you.
Current published rates, derived per-page costs, and how Claude compares with the rest of the market on the same basis.
You are deciding whether to call the Claude API directly for document extraction and need a number for the business case before you write the integration.
You process hundreds or thousands of invoices a month and want to know whether token billing or a flat plan is cheaper at your actual volume.
You are comparing Claude, GPT, Gemini and the dedicated document APIs and need every option expressed in the same unit: dollars per 1,000 invoice pages.
Anthropic does not sell OCR. There is no document endpoint, no prebuilt invoice model, and no per-page rate card. You send a PDF or an image to the ordinary Messages API and you are billed in tokens like any other request. That single fact explains most of the confusion around what Claude costs to read a document, and it is why every per-page number below is derived from published rates rather than quoted from a price list.
Last updated August 2026. All token rates in this section were re-fetched from Anthropic official model documentation on the date shown.
These are the published list prices per million tokens. Prompt caching and the Batch API both reduce them, and the Batch API discount is a flat 50 percent on all usage.
| Model | Input per 1M tokens | Output per 1M tokens | Context window | Max output |
|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $50.00 | 1M | 128k |
| Claude Opus 5 | $5.00 | $25.00 | 1M | 128k |
| Claude Opus 4.8 | $5.00 | $25.00 | 1M | 128k |
| Claude Sonnet 5 | $2.00 | $10.00 | 1M | 128k |
| Claude Sonnet 4.6 | $3.00 | $15.00 | 1M | 128k |
| Claude Haiku 4.5 | $1.00 | $5.00 | 200k | 64k |
Note that Claude Sonnet 5 is cheaper than the older Claude Sonnet 4.6 it replaces, at $2 and $10 rather than $3 and $15. If you are still calling Sonnet 4.6 for document work, moving to Sonnet 5 cuts the bill by a third with no other change.
Anthropic documents the image cost precisely: Claude views images in patches instead of pixels, each patch is a 28 by 28 pixel block called a visual token, and an image therefore costs ceil(width / 28) x ceil(height / 28) visual tokens.
A US Letter page scanned at 150 DPI is 1,275 by 1,650 pixels. That gives ceil(1275 / 28) = 46 and ceil(1650 / 28) = 59, so 2,714 visual tokens per page. We checked this formula against all four worked examples Anthropic publishes and it reproduces every one exactly, including the downscaled 4K case.
Two resolution tiers apply. Claude 4.7 and later models accept up to 2,576 pixels on the long edge and 4,784 visual tokens. Every earlier model caps at 1,568 pixels and 1,568 visual tokens, and images above either limit are downscaled before billing. A full invoice page exceeds the older cap, so on Claude Haiku 4.5 the image half of the bill is pinned at 1,568 tokens.
When you send a PDF, Anthropic converts each page to an image and extracts its text layer, and both are billed as input. The documentation puts the text side at 1,500 to 3,000 tokens per page depending on content density. Add the 2,714 visual tokens from above and a single PDF invoice page costs roughly 4,200 to 5,700 input tokens before your prompt.
There are no additional PDF fees on top of that. Standard token pricing is the whole story, which is genuinely simpler than the per-feature pricing on AWS or Azure. It is just not cheaper.
Assumptions, stated so you can change them: US Letter at 150 DPI (2,714 visual tokens), Anthropic published text range of 1,500 to 3,000 tokens per page, a 200 token extraction prompt, and 700 output tokens for a structured JSON result with line items. The low figure is a clean single page invoice, the high figure is a dense multi line one.
| Model | Per 1,000 pages | With Batch API |
|---|---|---|
| Claude Haiku 4.5 | $6.77 to $8.27 | $3.38 to $4.13 |
| Claude Sonnet 5 | $15.83 to $18.83 | $7.91 to $9.41 |
| Claude Sonnet 4.6 | $23.74 to $28.24 | $11.87 to $14.12 |
| Claude Opus 4.8 | $39.57 to $47.07 | $19.79 to $23.54 |
| Claude Opus 5 | $39.57 to $47.07 | $19.79 to $23.54 |
| Claude Fable 5 | $79.14 to $94.14 | $39.57 to $47.07 |
Rasterize the page and send it as a PNG or JPEG and the text layer disappears from the bill. You lose the embedded text, which matters on born-digital PDFs where that layer is perfectly accurate and free of OCR error, but on scans it is noise anyway.
| Model | Per 1,000 pages | With Batch API | Saving vs PDF |
|---|---|---|---|
| Claude Haiku 4.5 | $5.27 | $2.63 | 22% to 36% |
| Claude Sonnet 5 | $12.83 | $6.41 | 19% to 32% |
| Claude Sonnet 4.6 | $19.24 | $9.62 | 19% to 32% |
| Claude Opus 5 | $32.07 | $16.04 | 19% to 32% |
| Claude Fable 5 | $64.14 | $32.07 | 19% to 32% |
Every figure below is per 1,000 invoice pages on the same basis, drawn from the vendor pricing pages we maintain. Claude is the most expensive frontier reading layer in this market, and the reason is structural rather than a pricing choice.
| Option | Per 1,000 pages | Billing model |
|---|---|---|
| Gemini 2.5 Flash-Lite | $0.27 | Tokens, 258 per page |
| GPT-5 nano | $0.47 | Tokens, patch based |
| Gemini 2.5 Flash | $1.58 | Tokens, 258 per page |
| GPT-5 mini | $2.06 | Tokens, patch based |
| Mistral OCR 4 | $4.00 | Flat per page |
| Claude Haiku 4.5 | $5.27 | Tokens, 28px patches |
| Gemini 2.5 Pro | $6.32 | Tokens, 258 per page |
| GPT-4.1 | $7.53 | Tokens, tile based |
| GPT-5 | $9.12 | Tokens, patch based |
| AWS Textract Analyze Expense | $10.00 | Flat per page |
| Azure prebuilt invoice | $10.00 | Flat per page |
| Claude Sonnet 5 | $12.83 | Tokens, 28px patches |
| Claude Opus 5 | $32.07 | Tokens, 28px patches |
The structural reason Claude costs more: image tokenization. Google counts a document page as a flat 258 tokens regardless of its pixel dimensions. Claude counts the same page as 2,714 visual tokens, because it tokenizes on a 28 pixel grid rather than assigning a fixed per-page cost. That is roughly ten and a half times more image tokens for identical input, and it is the single largest driver of the gap. It is a deliberate design tradeoff in favor of fine visual detail, not an oversight, and it is why Claude does well on dense tables and degraded scans.
| Limit | Value |
|---|---|
| Maximum request size | 32 MB, lower on some partner platforms |
| Maximum pages per request | 600, or 100 when the context window is under 1M tokens |
| PDF format | Standard PDF only, no passwords or encryption |
| Image formats | JPEG, PNG, GIF, WebP |
| Maximum image dimensions | 8000 x 8000 pixels |
| Maximum image size | 10 MB base64 on the Claude API, 5 MB on Bedrock and Google Cloud |
| Images per request | 600, or 100 on 200k context models |
| Batch API discount | 50 percent on all token usage |
Two limits catch teams out. Claude Haiku 4.5, the cheapest option, has a 200k context window, so it is capped at 100 pages per request rather than 600. And if a single request contains more than 20 images a stricter per-image dimension limit applies, so resize to 2,000 pixels or keep batches at 20 documents or fewer.
No. Anthropic does not sell an OCR product or a document extraction endpoint. Claude reads PDFs and images through the standard Messages API using its vision capability, billed in tokens at the normal model rate. There is no OCR SKU, no prebuilt invoice model, and no per-page price to quote.
There is no official per-page price, but you can derive one. A US Letter invoice page sent as a PDF costs roughly $0.0068 to $0.0083 on Claude Haiku 4.5, about $0.016 to $0.019 on Claude Sonnet 5, and about $0.040 to $0.047 on Claude Opus 5, including a structured JSON response. Batch processing halves all of those.
Claude reads images in 28 by 28 pixel patches, each counted as one visual token, so an image costs ceil(width / 28) x ceil(height / 28) tokens. A 1,275 by 1,650 pixel page is 46 x 59, or 2,714 visual tokens. Images beyond the model resolution tier are downscaled first, which caps the token count.
Only at the Haiku tier. Claude Haiku 4.5 reads an invoice page for about $5.27 per 1,000 against $10.00 for Textract Analyze Expense, so it is roughly half. Claude Sonnet 5 at $12.83 and Claude Opus 5 at $32.07 are both more expensive than Textract, and neither returns the structured invoice schema Textract does out of the box.
Image tokenization. Gemini counts a document page as a flat 258 tokens no matter its dimensions, while Claude tokenizes on a 28 pixel grid and charges 2,714 visual tokens for the same page. That is about ten and a half times more image tokens, and it dominates the cost difference even before the per-token rates diverge.
Yes. A PDF page is billed for both its extracted text layer, which Anthropic puts at 1,500 to 3,000 tokens, and the rendered page image. An image is billed for the image alone. Rasterizing pages before sending cuts roughly 19 to 36 percent off the bill, at the cost of losing the embedded text layer.
No. Anthropic states that standard API pricing applies with no additional PDF fees. You pay only the token cost of the text and images the pages produce, which is simpler than the per-feature add-on pricing used by AWS Textract and Azure Document Intelligence.
Use Claude Haiku 4.5, send rasterized images rather than PDFs, scan at 150 DPI rather than 300, keep the output schema tight, and submit through the Batch API. Together those take a page from about $0.047 on Opus 5 to about $0.0026, a reduction of roughly 94 percent, in exchange for latency and some accuracy on difficult documents.
No. Anthropic does not sell an OCR product or a document extraction endpoint. Claude reads PDFs and images through the standard Messages API using its vision capability, billed in tokens at the normal model rate. There is no OCR SKU, no prebuilt invoice model, and no per-page price to quote.
There is no official per-page price, but you can derive one. A US Letter invoice page sent as a PDF costs roughly $0.0068 to $0.0083 on Claude Haiku 4.5, about $0.016 to $0.019 on Claude Sonnet 5, and about $0.040 to $0.047 on Claude Opus 5, including a structured JSON response. Batch processing halves all of those.
Claude reads images in 28 by 28 pixel patches, each counted as one visual token, so an image costs ceil(width / 28) x ceil(height / 28) tokens. A 1,275 by 1,650 pixel page is 46 x 59, or 2,714 visual tokens. Images beyond the model resolution tier are downscaled first, which caps the token count.
Only at the Haiku tier. Claude Haiku 4.5 reads an invoice page for about $5.27 per 1,000 against $10.00 for Textract Analyze Expense, so it is roughly half. Claude Sonnet 5 at $12.83 and Claude Opus 5 at $32.07 are both more expensive than Textract, and neither returns the structured invoice schema Textract does out of the box.
Image tokenization. Gemini counts a document page as a flat 258 tokens no matter its dimensions, while Claude tokenizes on a 28 pixel grid and charges 2,714 visual tokens for the same page. That is about ten and a half times more image tokens, and it dominates the cost difference even before the per-token rates diverge.
Yes. A PDF page is billed for both its extracted text layer, which Anthropic puts at 1,500 to 3,000 tokens, and the rendered page image. An image is billed for the image alone. Rasterizing pages before sending cuts roughly 19 to 36 percent off the bill, at the cost of losing the embedded text layer.
No. Anthropic states that standard API pricing applies with no additional PDF fees. You pay only the token cost of the text and images the pages produce, which is simpler than the per-feature add-on pricing used by AWS Textract and Azure Document Intelligence.
Use Claude Haiku 4.5, send rasterized images rather than PDFs, scan at 150 DPI rather than 300, keep the output schema tight, and submit through the Batch API. Together those take a page from about $0.047 on Opus 5 to about $0.0026, a reduction of roughly 94 percent, in exchange for latency and some accuracy on difficult documents.
The other frontier lab with no OCR product: full GPT-4o and GPT-5 per-page math, from $0.47 per 1,000.
The cheapest token-metered option at $0.27 per 1,000 pages, and the 258-token-per-page rule behind it.
A flat $4 per 1,000 pages with no token math at all. See how flat and metered billing compare.
Analyze Expense is $10 per 1,000 pages and returns a structured invoice schema out of the box.
The prebuilt invoice model, also $10 per 1,000 pages, with the strongest line item scores we have seen.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
billed as
$288 yearly
Choose speed vs accuracy when extracting
| Base AI Faster | 2,500 pages |
| Pro AI Best accuracy | 500 pages |
Scale invoice extraction across your whole team with automation.
USD
per month
billed as
$888 yearly
Choose speed vs accuracy when extracting
| Base AI Faster | 10,000 pages |
| Pro AI Best accuracy | 2,000 pages |
Enterprise‑grade invoice extraction, security, and controls.
USD
per month
billed as
$ yearly
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |
