Which LLM Is Cheapest for OCR? Every Model, Per 1,000 Pages

Gemini 2.5 Flash-Lite is the cheapest language model for OCR, at roughly $0.33 per 1,000 pages. OpenAI's GPT-6 Luna is next at about $0.62, or about $0.31 through the Batch API, and GPT-5.4-nano follows at about $1.44. After that the field jumps: GPT-5.4-mini and Claude Haiku 4.5 both land around $5, GPT-6 Sol and Claude Sonnet 5 sit near $12.50, and the flagship models run from $25 to $63 per 1,000 pages.

None of those numbers are published by anyone. Every vendor sells tokens, so each figure has to be derived from token prices plus that vendor's own rules for turning a page image into tokens. Below is the whole field in one place, converted on the same assumptions from each vendor's primary documentation. The OpenAI rows were recalculated in September 2026 after OpenAI released GPT-6 and published a single 1.2 image multiplier for its current models.

Every model, cheapest first

All figures assume one scanned US Letter page at 150 DPI, which is 1,275 by 1,650 pixels, transcribed to text at roughly 750 output tokens.

That is a spread of nearly 200 times between the cheapest and the dearest sensible option, for the same task. The full derivation per vendor lives on the LLM OCR pricing pillar, with the working for each on OpenAI, Claude and Gemini.

Why Gemini wins on price

Google charges a flat 258 tokens per page regardless of how large the page is. OpenAI and Anthropic both scale the input with page size until they hit a cap, so a letter page costs OpenAI 2,080 patches and Anthropic 2,714 visual tokens. Gemini charges 258 either way.

That makes the Gemini input side almost free, which is exactly why output tokens are 92 to 96 percent of a real Gemini OCR bill. It is also why you will see Gemini OCR quoted around three cents per 1,000 pages in various guides. That figure counts input only. The real number at a realistic 750 output tokens per page is about $0.33, roughly thirteen times higher, and still the cheapest option available.

The cheapest Claude models see less of the page

Here is the part that does not appear on any price list, and the reason the cheapest row is not automatically the right answer.

Anthropic caps its standard tier at 1,568 visual tokens against 4,784 on the high resolution tier, shrinking a letter page from 2,714 to about 1,560 tokens, a 42 percent cut. OpenAI used to do something similar with GPT-5.4-mini and nano, but its current vision guide gives them the same 2,500 patch budget and the same 1.2 multiplier as GPT-5.5 and GPT-6 Astra, so on OpenAI a 150 DPI letter page is 2,496 input tokens whichever model reads it.

On Claude, then, part of what you save on a budget model is resolution, and resolution is precisely what dense tables, small print and poor scans need. On Claude, raising your scan DPI does not help, because the extra pixels are scaled straight back out. On OpenAI it costs you: GPT-5.5 and GPT-6 Astra read a 300 DPI page at original detail by default, which roughly doubles the per-page bill.

Practical version: test the cheap model on your worst documents, not your best ones. A clean typed invoice will come back fine on nano. A carbon copy bill of lading with a six point font in the footer is where the resolution cut shows up, and it shows up as a wrong value rather than an error.

Cheapest is not the same as cheapest per correct page

Every model in that table shares three properties that a dedicated OCR engine does not have.

There is no per-value confidence score. Azure Read, AWS Textract and Google Enterprise Document OCR all tell you they read a digit at 62 percent confidence, which is what lets you route that one field to a human and leave the other forty alone. A language model hands you a value with no score attached. You either check everything or you accept silent errors.

There is no determinism. The same page can come back slightly differently on two runs. For a reconciliation process that expects stable output, that is a design constraint, not a detail.

And there is the failure mode that costs the most: when a model cannot read a total, it can produce a plausible number in the right format rather than failing. A classical engine returns garbage or a low score, which is easy to catch automatically. On financial documents this difference matters more than any per-page rate.

Retries belong in the budget too. A malformed JSON response or a truncated transcription means sending the page again at full price. Assume some percentage and price it in.

How this compares to dedicated OCR

AWS Textract, Azure AI Document Intelligence Read and Google Enterprise Document OCR all charge $1.50 per 1,000 pages for plain text extraction, dropping to $0.60 at volume on the tiered vendors. Azure AI Content Understanding Basic is $1.00.

So Gemini 2.5 Flash-Lite, GPT-6 Luna and GPT-5.4-nano are genuinely below that line, and everything from GPT-5.4-mini upward is above it. This is a real change. For years the argument against using a language model for OCR was cost, and it was a good argument. At the bottom of the range that argument is finished. The confidence score and determinism arguments are not.

Which one should you actually use

For high volume, clean, predictable pages, Gemini 2.5 Flash-Lite or batched GPT-6 Luna. Keep the output short and structured, because output is where nearly all the money goes.

For structured field extraction at volume, GPT-5.4-mini or Claude Haiku 4.5 at around $5 per 1,000 pages. Both return a JSON schema reliably. Watch dense tables on Haiku, since it sits on Anthropic's standard resolution tier. Mistral Document AI at $5.00 is worth testing as well, because it is a flat published page rate with no token modelling to do.

For messy scans, handwriting and unusual layouts, Claude Sonnet 5 at about $12.93 or GPT-6 Sol at about $12.49. Both read a 150 DPI letter page without shrinking it. This is the workload language models were genuinely worth a premium for, and the right comparison there is not $1.50 per 1,000 pages, it is the cost of a person keying it.

For financial and legal documents, route rather than choose. Use a classical engine for the confidence scores that tell you what to review, and a model for the fields the engine cannot structure.

The cost nobody puts in the comparison

Whichever row you pick, the extraction call is the cheap part. The validation rules, the human review queue, the retry handling and the export are the work. So is what happens downstream: extracted fields usually land in a warehouse table that somebody now depends on, and a silent change in how a model formats a date can quietly corrupt a month of records before anyone notices, which is why teams running these pipelines at scale end up watching those tables for schema drift and volume anomalies rather than trusting the extraction step blindly.

That is the honest way to compare a full product like DocuOCR at roughly $14 to $20 per 1,000 pages. Not against $0.33, but against $0.33 plus everything above, built and maintained by you. Sometimes building it is right. Upload a document and look at the output before you decide.