Gemini OCR Pricing per Page and per Receipt

Gemini OCR pricing is per token, not per page. On the paid Gemini API a receipt photo costs roughly $0.50 to $3.40 per 1,000 receipts on the Flash and Flash-Lite models at standard rates, and half that through the Batch API. Gemini 3.8 Flash input is $0.75 per million tokens through December 31, 2026 and doubles to $1.50 on January 1, 2027. Google Document AI's Expense parser, by contrast, is a flat $0.10 per receipt.

Every figure below was read from Google's Gemini API pricing page and its media resolution and document processing docs in October 2026. Google publishes no per page price for Gemini, so the per receipt costs are our arithmetic from those token rates, with every assumption shown. If you are pricing an extraction feature for a product, this is the budget view. If you just need receipts in a spreadsheet or QuickBooks, skip to the last section, because the cheapest token is the one you never have to engineer around.

Gemini API prices for OCR models per 1 million tokens

Images and PDF pages are billed as input tokens at the text rate. Google states that tokens for the document modality, PDFs included, are billed at the image token rate. Output is the JSON or text the model returns, and on models that think, the thinking tokens are billed as output too.

Gemini 3.6 Flash carries the same card as 3.8 Flash, including the January 2027 increase. Flex pricing matches Batch on every model above. Priority processing costs 1.8 times standard, which is $1.35 input and $6.75 output on 3.8 Flash in 2026. The free tier exists, but Google's terms say free tier content is used to improve its products, and paid tier content is not. For customer receipts and vendor invoices that one line settles the question: budget for the paid tier.

How many tokens does a receipt use in Gemini?

It depends on the model family and on one setting. On Gemini 3 models the media_resolution parameter caps the tokens per image or page. Google's table gives 1,120 tokens for an image at the default and high settings, 560 at medium and 280 at low. A PDF page defaults to 560 tokens, and Google recommends medium for PDFs, noting that OCR quality on standard documents typically saturates there.

Gemini 2.5 models use the older tiling rule. An image of 384 pixels or less on both sides is 258 tokens. Anything larger is cut into 768 by 768 tiles at 258 tokens each, and a tall receipt photo of about 1,000 by 2,400 pixels comes to 8 tiles, or 2,064 tokens. A PDF page on these models is 258 tokens regardless of size. That is why a PDF receipt is cheaper than a phone photo of the same receipt on 2.5 Flash.

Two other inputs add up. Your prompt, with the schema of fields you want back, is often 300 to 500 tokens. The answer itself, a JSON object with merchant, date, totals, tax and fifteen line items, lands around 500 to 700 tokens.

Gemini OCR cost per 1,000 receipts

Here is the arithmetic for a 1,000 receipt month. Assumptions: a 400 token prompt, 600 output tokens, photos at 1,120 tokens on Gemini 3 and 2,064 on 2.5, PDF pages at 560 on Gemini 3 and 258 on 2.5. The last column adds 1,000 thinking tokens per receipt, which is a modest amount for a reasoning model left on its default settings.

The thinking column is the one people miss. On 3.8 Flash, a thousand tokens of reasoning more than doubles the bill, because output costs five times what input does. If your extraction does not need reasoning, set the thinking budget low and measure. If it does, that is the real price, and it belongs in the plan.

Is Gemini cheaper than Google Document AI for receipts?

On raw API cost, by a wide margin. Google Document AI prices the Expense parser, which its price list says was formerly the receipt parser, at $0.10 per document, the same as the Invoice parser. That is $100 per 1,000 receipts, against $1 to $7 on the Gemini Flash models above. Enterprise Document OCR on Document AI is $1.50 per 1,000 pages after the first 1,000 free each month, but it returns text and layout, not fields.

The gap is real, and it is also the cost of the work you take on. The Expense parser returns typed entities with a fixed schema. With Gemini you write the schema, the prompt, the validation that totals add up, the retry when JSON comes back malformed, and the review step for low confidence fields. Our Google Document AI pricing page breaks down each processor, and the Google Cloud Vision API pricing article covers the plain text detection route.

Gemini vs other LLM OCR APIs on price

For the page priced LLM services, our Mistral OCR API pricing page has the current card, and the OCR API pricing hub compares the big clouds side by side.

What the token price leaves out

Budgets built on token prices alone tend to be wrong in the same three places.

First, retries. A model that returns invalid JSON or a total that does not match its line items gets called again. A 5% retry rate is a 5% higher bill, and on photos of crumpled gas station receipts the rate is rarely that low on the first prompt version.

Second, review. Somebody looks at the receipts where the model was unsure. If that is a bookkeeper at $30 an hour spending 20 seconds on one receipt in ten, the review costs about $17 per 1,000 receipts, several times the Gemini bill on any Flash model.

Third, the documents themselves. A vendor invoice is untrusted input going straight into a language model. Text hidden in a PDF can try to steer the model, and if your pipeline lets the model call tools or write to your books, that matters. Teams wiring an LLM into financial workflows should put prompt injection protection for AI agents in front of anything the model is allowed to act on.

When a flat plan beats building on Gemini

If you are a developer adding receipt capture to a product with thousands of users, Gemini Flash-Lite at about a dollar per 1,000 receipts is hard to beat, and you should build on it. Plan for the 2027 increase on 3.8 Flash and test 3.1 Flash-Lite first.

If you are an accountant, bookkeeper or operations lead and the goal is a clean spreadsheet or a QuickBooks import by Friday, the API price is the smallest number in the project. ReceiptOCR Starter is $49 a month for 2,500 pages, Plus is $149 for 10,000, and both include the prompt engineering, validation, review screen and exports you would otherwise build. You upload a batch on the bulk receipt scanner, check the flagged fields, and download Excel, CSV or QBO. Developers who still want an endpoint can compare our receipt OCR API on the Pro plan.

Gemini OCR pricing FAQ

How much does Gemini OCR cost per page?

Google does not sell Gemini by the page. On Gemini 3 models a PDF page counts as up to 560 input tokens at the default setting, so on 3.1 Flash-Lite the page itself costs about $0.00014 and the full request, with prompt and JSON answer, about $0.0011. Output tokens usually cost more than the page.

Is the Gemini API free for OCR?

There is a free tier with rate limits in Google AI Studio and the API, but Google's terms say content sent on the free tier is used to improve its products. For receipts and invoices containing customer or vendor data, use the paid tier, where content is not used that way.

How much does Gemini 2.5 Flash OCR cost?

Gemini 2.5 Flash is $0.30 per million input tokens and $2.50 per million output tokens, half that on Batch. A PDF page is 258 tokens on this model, so a typical receipt request comes to about $1.70 per 1,000 receipts at standard rates, before any thinking tokens.

Does the Gemini Batch API make OCR cheaper?

Yes, by 50% on every model listed here. Batch suits month end backlogs and nightly imports where results can wait. It does not suit a user taking a photo of a receipt in your app and expecting fields back in seconds.

Is Gemini 3.8 Flash getting more expensive?

Yes. Google's pricing page lists Gemini 3.8 Flash at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026, and $1.50 and $7.50 starting January 1, 2027. Gemini 3.6 Flash carries the same schedule.

What is the maximum PDF size Gemini accepts?

Gemini accepts PDFs up to 50 MB or 1,000 pages, for both inline data and Files API uploads. Larger pages are scaled down to a maximum of 3,072 by 3,072 pixels and smaller ones scaled up to 768 by 768, with no cost reduction for small pages.