How much does LLM OCR cost per 1,000 pages?

It ranges from about $0.33 to about $197 per 1,000 pages depending on the model. Gemini 2.5 Flash-Lite is roughly $0.33, GPT-5.4-nano about $1.67, Claude Haiku 4.5 about $5.31, GPT-5.4-mini about $5.19, Claude Sonnet 5 about $19.39 and Claude Opus 4.8 about $32.32. Dedicated OCR from AWS, Azure and Google is a flat $1.50 for plain text.

Which LLM is cheapest for OCR?

Gemini 2.5 Flash-Lite, at roughly $0.33 per 1,000 pages, is the cheapest page-to-text path published by any vendor, language model or otherwise. GPT-5.4-nano is next at about $1.67, or about $0.84 through the OpenAI Batch API. Both are below the $1.50 that AWS Textract, Azure Document Intelligence Read and Google Enterprise Document OCR charge for plain text extraction.

Is LLM OCR cheaper than traditional OCR?

It now can be, which was not true a year ago. The cheapest language model paths, Gemini 2.5 Flash-Lite at about $0.33 and GPT-5.4-nano at about $1.67 per 1,000 pages, sit at or below the $1.50 dedicated OCR rate. Mid-tier models do not: Claude Sonnet 5 at about $19.39 is roughly 13 times the dedicated rate. The answer depends entirely on which model you pick.

How do you convert token pricing to a per-page OCR rate?

Three numbers. First the image tokens for one page, which comes from each vendor patch formula: OpenAI uses 32 by 32 pixel patches, Anthropic uses 28 by 28. Second the output tokens for the transcription, roughly 750 for a full page. Third the per-million input and output prices. Multiply, add, and multiply by 1,000. State your page size and output length, because both change the answer.

Do the cheap models read the page at full resolution?

No, and this is the trap the price list hides. OpenAI caps GPT-5.4-mini and nano at 1,536 image patches against 2,500 or more on the full models, so a letter page is shrunk from 2,080 patches to 1,496. Anthropic caps standard-tier models at 1,568 visual tokens against 4,784, shrinking the same page from 2,714 to about 1,560. Part of the discount is resolution.

Does batch processing reduce LLM OCR costs?

Yes, and it is the largest discount that actually applies. OpenAI, Google and Mistral all offer 50 percent off through a batch or asynchronous endpoint in exchange for a turnaround window of up to 24 hours. Document backlogs are the ideal workload for it. Prompt caching, which looks bigger, rarely helps, because the page image is the part of the request that changes every call.

Is an LLM accurate enough to replace OCR?

It depends on what you do with the output. A language model reads handwriting, odd layouts and poor scans better than a classical engine, but it returns no per-value confidence score and can produce a fluent, plausible, wrong value on a field it could not read. A dedicated engine tells you it read a digit at 62 percent confidence so you can route it for review. Most production pipelines use both.

What does LLM OCR cost for 1 million pages?

At the cheapest end, about $330 on Gemini 2.5 Flash-Lite or about $1,670 on GPT-5.4-nano, falling to roughly $840 in batch mode. In the middle, about $5,190 on GPT-5.4-mini or $5,310 on Claude Haiku 4.5. At the top, about $32,300 on Claude Opus 4.8. Dedicated OCR at that volume is $1,500 and falls to $600 per million on the tiered vendors.

Which LLM is best for document data extraction?

GPT-5.4-mini and Claude Haiku 4.5 are the usual working answers, both around $5 per 1,000 pages, cheap enough for volume and strong enough for structured fields. Move up to Claude Sonnet 5 or GPT-5.4 when the document genuinely needs reasoning. Use Gemini 2.5 Flash-Lite or GPT-5.4-nano when you are transcribing clean, predictable pages at scale.

// Verified rate pillar, July 22, 2026

LLM OCR Pricing: Cost Per 1,000 Pages by Model

Every vendor prices language models in tokens, so nobody publishes what LLM OCR costs per page. Converted, it runs from about $0.33 per 1,000 pages on Gemini 2.5 Flash-Lite to about $197 on the pro models. Here is every model in one table, against the $1.50 that dedicated OCR charges.

OpenAI, Anthropic, Google and Mistral
Why published estimates are 2x to 13x low
The resolution trap in the cheap models
What each vendor patch formula does

See every model Extract a document free

Upload a document, no signup

PDF, JPG, PNG, BMP, HEIC, TIFF

Upload a document to extract

Extract documents →

Rates are the easy part. Whether the fields come out right on your documents is the part that decides the bill. Drop one in and look at the output.

SOC 2 Type II

256-bit encryption

US data handling

Fields, not just text

$0.33: cheapest LLM, per 1,000 pages
$1.50: what dedicated OCR charges
600x: spread from cheapest to dearest
56-96%: of every LLM bill is output tokens

// The short answer

What LLM OCR costs

Reading a document with a language model costs between about $0.33 and about $197 per 1,000 pages, depending entirely on which model you choose. Gemini 2.5 Flash-Lite is roughly $0.33, GPT-5.4-nano about $1.67, GPT-5.4-mini about $5.19, Claude Haiku 4.5 about $5.31, Claude Sonnet 5 about $19.39 and Claude Opus 4.8 about $32.32. Dedicated OCR from AWS, Azure and Google is a flat $1.50 per 1,000 pages for plain text. The headline is that in 2026 the cheapest language models finally crossed that line, so price is no longer the reason to pick a classical OCR engine. What has not changed is that a classical engine returns a confidence score for every word it reads, and a language model does not.

// Every model, converted to dollars

LLM OCR pricing per 1,000 pages

One scanned US Letter page at 150 DPI, which is 1,275 by 1,650 pixels, transcribed to text at roughly 750 output tokens. Both assumptions are stated so you can redo the math for your own documents. Sorted cheapest first within each tier.

Model	Vendor	Per 1,000 pages	Output share of bill	Note
Gemini 2.5 Flash-Lite	Google	$0.33	92% to 96%	Cheapest page-to-text path published anywhere
GPT-5.4-nano (batch)	OpenAI	$0.84	56%	Undercuts dedicated OCR, 24 hour turnaround
GPT-5.4-nano	OpenAI	$1.67	56%	First LLM to match dedicated OCR at list price
Mistral OCR 4	Mistral	$4.00	n/a	Flat per-page rate, no token math, batch halves it
Mistral Document AI	Mistral	$5.00	n/a	Cheapest published structured extraction rate
GPT-5.4-mini	OpenAI	$5.19	65%	The usual working choice for volume extraction
Claude Haiku 4.5	Anthropic	$5.31	71%	Standard resolution tier, downscales dense pages
Claude Sonnet 4.6	Anthropic	$15.93	71%	Standard resolution tier
GPT-5.4 / GPT-5.6-terra	OpenAI	$16.45	68%	Input column estimated, no published multiplier
Claude Sonnet 5	Anthropic	$19.39	58%	$12.93 on the introductory rate to Aug 31, 2026
Claude Opus 4.8	Anthropic	$32.32	58%	Premium reasoning tier
GPT-5.5 / GPT-5.6-sol	OpenAI	$32.90	68%	Input column estimated
Claude Fable 5	Anthropic	$64.64	58%	The most expensive Claude path
GPT-5.5-pro / GPT-5.4-pro	OpenAI	$197.40	68%	Not a document processing price point

Derived July 22, 2026 from each vendor's own published token prices and image documentation. Full working for each vendor is on OpenAI OCR pricing, Claude OCR pricing, Gemini OCR pricing and Mistral OCR pricing. Re-verify before quoting.

// The line everything is measured against

What dedicated OCR costs

The three hyperscalers converged on the same number years ago and have not moved off it. This is the baseline every model in the table above is competing with.

Service	Per 1,000 pages	What you get	At volume
Google Document AI (Enterprise OCR)	$1.50	Text, coordinates, per-word confidence	Falls to $0.60 above 5M pages
AWS Textract (Detect Document Text)	$1.50	Text, coordinates, per-word confidence	Falls to $0.60 above 1M pages
Azure AI Document Intelligence (Read)	$1.50	Text, coordinates, per-word confidence	Commitment tiers down to $0.45
Azure AI Content Understanding (Basic)	$1.00	Text extraction	The cheapest published cloud OCR rate
Google Cloud Vision (text detection)	$1.50	Text and bounding boxes	Falls to $0.60 above 5M units

Note the third column. Every one of these returns a per-word confidence score. No language model in the table above does, and that gap is the reason plenty of teams still route financial and legal documents through a classical engine even when a model is cheaper.

// What the price lists do not say

Three things that hold across every vendor

We derived each vendor separately, from its own documentation. The same three patterns turned up every time.

Output tokens are the bill

Transcription is an output-heavy job, and every vendor charges five to eight times more for output than input. Output is 56 to 68 percent of an OpenAI bill, 58 to 71 percent of a Claude bill and 92 to 96 percent of a Gemini bill. Almost every figure published elsewhere counts input only, which is why you will see Gemini OCR quoted at three cents per 1,000 pages when the real number is about thirty-three. If you need six fields off an invoice, ask for six fields, not a transcription. Output length is the one variable you fully control, and on Gemini it is nearly the whole invoice.

The cheap models see less

Both OpenAI and Anthropic cap their budget models at a lower image resolution than their flagships, and neither puts that on the price list. OpenAI allows GPT-5.4-mini and nano 1,536 patches against 2,500 or more on the full models, shrinking a letter page by 28 percent. Anthropic allows standard-tier models 1,568 visual tokens against 4,784, shrinking the same page by 42 percent. Two companies arrived at the same design independently. Part of what you save is resolution, and resolution is exactly what dense tables and small print need. Scanning at higher DPI does not fix it, because the extra pixels are scaled straight back out.

2026 is the year price stopped mattering

For years the argument against language models for OCR was cost, and it was a good argument. That argument is now over at the bottom of the range. Gemini 2.5 Flash-Lite at about $0.33 and GPT-5.4-nano at about $1.67, or about $0.84 through batch, are at or under the $1.50 dedicated rate. What has not changed is the engineering around the model: no per-value confidence score, no guarantee the same page returns the same answer twice, and a real chance of a fluent wrong value on a field the model could not read. Those are now the reasons to choose, not the rate.

// How each vendor counts a page

The image math, vendor by vendor

Every vendor turns a page image into tokens differently, and the differences are large enough to flip a comparison. All figures below are for the same scanned US Letter page at 150 DPI.

Vendor and tier	Image unit	One letter page	Budget cap	What happens to the page
OpenAI (full models)	32 x 32 pixel patch	2,080 patches	2,500 at high detail	Nothing, a letter page fits
OpenAI (mini and nano)	32 x 32 pixel patch	1,496 patches	1,536 patches	Shrinks 28%, then a 1.62x or 2.46x multiplier
Anthropic (high-res tier)	28 x 28 pixel patch	2,714 visual tokens	4,784 tokens	Nothing, a letter page fits
Anthropic (standard tier)	28 x 28 pixel patch	~1,560 visual tokens	1,568 tokens	Shrinks 42%
Google Gemini	Fixed per page	258 tokens	n/a	Nothing, the rate is flat per page
Mistral	Priced per page	n/a	n/a	No token math, a flat published page rate

One consequence worth noting: Gemini charges a flat 258 tokens per page regardless of how big the page is, so its input cost barely moves, which is exactly why output ends up being 92 to 96 percent of a Gemini bill. OpenAI and Anthropic both scale input with page size until they hit a cap.

// Picking one

Which one to use, by workload

High volume, clean, predictable pages

Gemini 2.5 Flash-Lite or GPT-5.4-nano, batched. At roughly $0.33 and $0.84 per 1,000 pages these are cheaper than dedicated OCR, and on a clean typed page the accuracy gap is small. Keep the output short and structured, because output is where the money goes. This is the workload where the 2026 price shift genuinely changes your architecture.

Structured field extraction at volume

GPT-5.4-mini or Claude Haiku 4.5, both around $5 per 1,000 pages. Both are strong enough to return a JSON schema reliably and cheap enough to run on everything. Watch the resolution caps on dense tables, since both sit on their vendor budget tier. Mistral Document AI at $5.00 is worth testing too, because it is a flat per-page rate with no token math to model.

Messy scans, handwriting, unusual layouts

Claude Sonnet 5 at about $19.39, or GPT-5.4 at about $16.45. This is the workload language models were actually worth the premium for, and both sit on the full resolution tier so they see the whole page. Expensive per page, but on documents a template-based engine simply cannot read, the comparison is not against $1.50, it is against manual keying.

Financial and legal documents

Route them, do not pick one. Use a classical engine for the per-word confidence scores so you know which values to review, and a model for the fields the engine cannot structure. Anything where a wrong number carries real cost needs a validation layer and a human queue regardless of which model reads the page, and that layer is more work than the extraction call.

Whichever row you land on, the extraction call is the cheap part. The validation rules, the review queue, the retry handling and the export are what take the months. DocuOCR ships that assembled, at roughly $14 to $20 per 1,000 pages depending on plan, which is the honest way to compare it: not against a raw token rate, but against a raw token rate plus the pipeline you would otherwise write.

Frequently asked questions

How much does LLM OCR cost per 1,000 pages?: It ranges from about $0.33 to about $197 per 1,000 pages depending on the model. Gemini 2.5 Flash-Lite is roughly $0.33, GPT-5.4-nano about $1.67, Claude Haiku 4.5 about $5.31, GPT-5.4-mini about $5.19, Claude Sonnet 5 about $19.39 and Claude Opus 4.8 about $32.32. Dedicated OCR from AWS, Azure and Google is a flat $1.50 for plain text.
Which LLM is cheapest for OCR?: Gemini 2.5 Flash-Lite, at roughly $0.33 per 1,000 pages, is the cheapest page-to-text path published by any vendor, language model or otherwise. GPT-5.4-nano is next at about $1.67, or about $0.84 through the OpenAI Batch API. Both are below the $1.50 that AWS Textract, Azure Document Intelligence Read and Google Enterprise Document OCR charge for plain text extraction.
Is LLM OCR cheaper than traditional OCR?: It now can be, which was not true a year ago. The cheapest language model paths, Gemini 2.5 Flash-Lite at about $0.33 and GPT-5.4-nano at about $1.67 per 1,000 pages, sit at or below the $1.50 dedicated OCR rate. Mid-tier models do not: Claude Sonnet 5 at about $19.39 is roughly 13 times the dedicated rate. The answer depends entirely on which model you pick.
Why are published LLM OCR cost estimates so low?: Because almost all of them count input tokens only. Transcribing a page produces roughly 750 output tokens, billed at five to eight times the input rate, and that output is 56 to 68 percent of an OpenAI bill, 58 to 71 percent of a Claude bill and 92 to 96 percent of a Gemini bill. An input-only estimate understates the real cost by roughly two times on OpenAI and thirteen times on Gemini.
How do you convert token pricing to a per-page OCR rate?: Three numbers. First the image tokens for one page, which comes from each vendor patch formula: OpenAI uses 32 by 32 pixel patches, Anthropic uses 28 by 28. Second the output tokens for the transcription, roughly 750 for a full page. Third the per-million input and output prices. Multiply, add, and multiply by 1,000. State your page size and output length, because both change the answer.
Do the cheap models read the page at full resolution?: No, and this is the trap the price list hides. OpenAI caps GPT-5.4-mini and nano at 1,536 image patches against 2,500 or more on the full models, so a letter page is shrunk from 2,080 patches to 1,496. Anthropic caps standard-tier models at 1,568 visual tokens against 4,784, shrinking the same page from 2,714 to about 1,560. Part of the discount is resolution.
Does batch processing reduce LLM OCR costs?: Yes, and it is the largest discount that actually applies. OpenAI, Google and Mistral all offer 50 percent off through a batch or asynchronous endpoint in exchange for a turnaround window of up to 24 hours. Document backlogs are the ideal workload for it. Prompt caching, which looks bigger, rarely helps, because the page image is the part of the request that changes every call.
Is an LLM accurate enough to replace OCR?: It depends on what you do with the output. A language model reads handwriting, odd layouts and poor scans better than a classical engine, but it returns no per-value confidence score and can produce a fluent, plausible, wrong value on a field it could not read. A dedicated engine tells you it read a digit at 62 percent confidence so you can route it for review. Most production pipelines use both.
What does LLM OCR cost for 1 million pages?: At the cheapest end, about $330 on Gemini 2.5 Flash-Lite or about $1,670 on GPT-5.4-nano, falling to roughly $840 in batch mode. In the middle, about $5,190 on GPT-5.4-mini or $5,310 on Claude Haiku 4.5. At the top, about $32,300 on Claude Opus 4.8. Dedicated OCR at that volume is $1,500 and falls to $600 per million on the tiered vendors.
Which LLM is best for document data extraction?: GPT-5.4-mini and Claude Haiku 4.5 are the usual working answers, both around $5 per 1,000 pages, cheap enough for volume and strong enough for structured fields. Move up to Claude Sonnet 5 or GPT-5.4 when the document genuinely needs reasoning. Use Gemini 2.5 Flash-Lite or GPT-5.4-nano when you are transcribing clean, predictable pages at scale.

See what the output actually looks like

Rates are easy to compare and tell you almost nothing. Upload a document to DocuOCR, look at the extracted fields, and decide from there.

Extract a document free Compare every OCR rate

Vendor rate references

OpenAI OCR pricing

GPT-5.4-nano $1.67, mini $5.19, GPT-5.5 $32.90

Claude OCR pricing

Haiku $5.31, Sonnet 5 $19.39, Opus $32.32

Gemini OCR pricing

Why output is 92 to 96 percent of the bill

Mistral OCR pricing

$4 per 1,000 pages, and half that in batch

LLM OCR

When a language model beats classical OCR

OCR pricing per 1,000 pages

Every vendor, language model or not

Agentic document extraction pricing

LlamaParse, Reducto, Landing AI, Extend

OCR API pricing comparison

The buyer view across the market

OCR API limits comparison

Page, size and rate ceilings by vendor

Self hosted OCR cost

What a GPU costs when you skip the API entirely

olmOCR pricing

The open model, sourced and repriced on AWS

From the same family of tools