Gemini OCR Pricing 2026: Gemini OCR API Cost Per 1,000 Pages and Flash OCR Rates

Gemini OCR pricing: official token rates, August 2026

Google does not publish a per-page OCR rate. The Gemini API bills tokens, and document pages enter as image tokens. These are the current published rates on the paid Standard tier, taken from Google’s developer pricing page in July 2026. Output prices include thinking tokens.

Two rates carry a surcharge on long prompts. Gemini 2.5 Pro moves to $2.50 input and $15.00 output above 200,000 tokens, and Gemini 3.1 Pro Preview moves to $4.00 input and $18.00 output. For invoice work you almost never cross that line, because a single invoice page is 258 tokens.

Gemini OCR cost per 1,000 pages, worked out

Google states that each document page is equivalent to 258 tokens, and that tokens from text natively embedded in a PDF are not charged. So 1,000 invoice pages is 258,000 input tokens. Output is the part you control: a plain text transcription is short, while a structured invoice record with header fields and line items is longer. The table below assumes 600 output tokens per page, which is a realistic figure for an invoice JSON record with a dozen header fields and ten line items. Change that assumption and the output column scales linearly.

The headline is that the reading step is close to free. On Gemini 2.5 Flash-Lite, the input cost of putting 1,000 invoice pages in front of the model is about three cents. Even on the most expensive Pro preview model, a thousand pages of raw input is around fifty cents. Output tokens, not page tokens, are what your Gemini OCR bill is actually made of.

One caveat that catches people out on their first invoice: the 2.5 and 3 series are thinking models, and Google bills thinking tokens at the output rate. If you leave a high thinking budget on while parsing a table-heavy invoice, the model can spend several hundred reasoning tokens before it writes a single field, and your real output volume lands well above the 600 assumed here. Cap the thinking budget for extraction work and the numbers above hold.

Gemini OCR pricing vs the per-page OCR APIs

Every other name in this market bills a flat rate per page, so comparing them takes one more step. Here is the same 1,000 pages priced across the options, with the Gemini figure using 2.5 Flash-Lite at the assumption above.

Read that table honestly and it says something uncomfortable for everyone selling per-page OCR: on raw model cost, Gemini undercuts the dedicated OCR APIs by an order of magnitude. It also says something the price column hides. The first three rows are a model endpoint. The last four are a product. What sits between them is the schema design, the prompt engineering, the retry and validation logic, the confidence handling, the storage, the review screen for the fields the model got wrong, and the export your bookkeeper actually opens. None of that is in the $0.27.

What the Gemini free tier really costs you

Gemini Flash and Flash-Lite models have a free tier, and that is why so many people search for free Gemini OCR. For a personal project it is genuinely free. For invoice work at a US business it usually is not usable, for one reason that has nothing to do with rate limits: Google’s own pricing page marks the free tier as content used to improve our products, and the paid tier as content not used to improve our products. Client invoices carry vendor names, bank details, addresses and payment terms. If you handle other people’s financial documents, the paid tier is the only defensible choice, and at three cents per thousand pages of input it is not the part of the budget worth optimizing.

The other free tier constraint is throughput. Free tier limits are set per project and Google publishes them in AI Studio rather than in the docs, and they are explicitly not guaranteed. Paid usage tiers move in steps: Tier 1 starts once billing is active with a $250 spend cap, Tier 2 opens at $100 spent plus three days, and Tier 3 at $1,000 spent plus thirty days. A month-end AP run that tries to push several thousand invoices through a free tier key will hit a wall.

The limits that shape a Gemini invoice pipeline

• 1,000 pages maximum per document, and up to 1,000 combined pages across multiple PDFs in one request.
• 50 MB maximum file size. Scanned invoice batches from an office multifunction device cross this more often than you would think.
• Pages are scaled to a maximum of 3,072 by 3,072 pixels and a minimum of 768 by 768, so very high resolution scans gain you nothing and very low resolution scans are upsampled rather than improved.
• Use the Files API rather than inline data for larger documents and anything you will send more than once. It cuts request latency and bandwidth.
• PDF is the only document format that gets real vision processing. TXT, Markdown, HTML and XML are accepted but read as plain text, so anything visual in them is lost.
• Batch and caching discounts exist. The Batch API is a 50 percent cost reduction, and context caching is billed separately per hour of storage.

Is Gemini OCR better than Mistral OCR?

On current independent testing, yes on accuracy and clearly yes on price. OCR Arena runs blind community voting between OCR models on real documents, and its published record puts Gemini 3 Flash at ELO 1821 and rank #1 against Mistral OCR v3 at 1576 and rank #9, with Gemini taking 20 of 23 head to head matchups. On cost, 1,000 invoice pages runs about $0.27 on Gemini 2.5 Flash-Lite against a flat $4 per 1,000 for Mistral OCR 4.

Mistral still wins several rows that often decide a purchase. It returns paragraph level bounding boxes and per block confidence scores, which Gemini does not document at all, and those matter the moment a human has to approve an extraction against the original page. Its flat per page rate is forecastable in a way a token bill is not. And it can be self hosted as a single container, while Gemini has no on premise option, so if your controls do not allow vendor documents to leave your network then Gemini is not a candidate at any price.

One honest caveat on the benchmark: it measures Mistral OCR v3, not the newer Mistral OCR 4 released in June 2026, and community voting rewards clean transcription rather than whether the fourth row of a continuation table survived. For accounts payable that second thing is what actually costs money. The full side by side, including a per model cost table, is on our Gemini OCR vs Mistral OCR comparison.

When Gemini OCR pricing beats a flat invoice tool, and when it does not

If you have engineers, a queue, somewhere to put the data, and volume in the hundreds of thousands of pages, Gemini is the cheapest reading layer available and the build is worth doing. The token cost genuinely rounds to nothing at that scale, and you keep full control of the schema.

If you are a controller, a bookkeeper, or a three-person finance team, the calculation inverts. A developer week at US rates costs more than five years of our Plus plan. You are not really choosing between $0.27 and $14.90 per thousand pages; you are choosing between a project and a login. Teams that already meter a lot of API spend across their stack tend to find that the invoice line is trivial next to everything else, which is its own argument for not building it: the money is in engineering hours, not tokens, and keeping a clear view of cloud and SaaS spend usually shows the same pattern across the rest of the stack.

Our own pricing is flat on purpose. Starter is $49 a month for 2,500 Base AI pages, which is about $19.60 per thousand. Plus is $149 a month for 10,000 Base AI pages, about $14.90 per thousand. Nobody watches a token meter, nobody writes a schema, and the output lands in Excel, CSV, JSON or QBO. That is a worse per-page rate than Gemini and a better total cost for most finance teams, and both of those statements are true at the same time.

The three newest Flash models are half price until 31 December 2026

Gemini 3.8 Flash, Gemini 3.7 Flash and Gemini 3.6 Flash all carry the same promotional rate on Google's pricing page: $0.75 per million input tokens and $3.75 per million output tokens through 31 December 2026, rising to $1.50 and $7.50 after that date. The table above uses the post-promotional rates, because those are the numbers a budget written today should survive on. If you are pricing a project that ships this year, the real cost is half.

Digital PDFs cost less than scans on Gemini 3 models

Google's document processing documentation states that each document page is equivalent to 258 tokens, and that on Gemini 3 models the tokens from text natively embedded in a PDF are not charged, while pages processed as images are counted under the image modality. In practice that means a born-digital vendor invoice with a real text layer bills less than a scanned one, and the difference is not something the API tells you about after the fact. If most of your AP volume arrives as emailed digital PDFs rather than scans, your effective Gemini rate is lower than any table can show. Batch requests take a further 50% off across the models that support them, and a single request accepts up to 1,000 pages or 50 MB.

Rates and limits on this page re-read from Google's official Gemini API pricing and document processing documentation on 26 August 2026.