Google publishes no per-page OCR rate for Gemini. It publishes two numbers that produce one: each document page is 258 tokens, and tokens are billed per million. Do the arithmetic and Gemini OCR lands at about $0.33 per 1,000 pages on Gemini 2.5 Flash-Lite, about $1.95 on Gemini 2.5 Flash, and about $3.23 on the newest Gemini 3.8 Flash. On Gemini 3 models a PDF page is 560 tokens by default, not 258.
The catch nearly every guide misses: output tokens, not the pages, are 87% to 96% of the bill.
Upload a document to extract
Drop files here or click to upload
Up to 50 files
The free demo extracts the first 5 files, the rest unlock when you upgrade
Uploading...
Before you estimate a Gemini token bill, drop in the document you meant to test on it and see what finished, validated fields look like.
Gemini has no OCR price list. It has a token price list, plus one line in the document-processing docs stating that each document page is equivalent to 258 tokens. Put those together and 1,000 pages costs 258,000 input tokens, which on the cheapest model is about three cents. That is the number most guides publish, and it is wrong, because OCR does not just read a page, it writes the page back out. Those output tokens are billed at four to eight times the input rate on every Gemini model, and a dense page emits far more tokens than it consumes. Counting both sides, Gemini 2.5 Flash-Lite lands near $0.33 per 1,000 pages, Gemini 2.5 Flash near $1.95, Gemini 3.8 Flash near $3.23 and Gemini 3.5 Flash near $7.59. The Gemini 3 figures are higher partly because those models bill a PDF page at 560 tokens by default, more than twice the 258 that every guide quotes. Even the corrected Flash-Lite figure is remarkable: roughly four and a half times cheaper than Azure Read, AWS Textract, and Google's own Document AI, all of which sit around $1.50 per 1,000 pages. Gemini really is the cheapest way to turn a page into text. It is not the cheapest way to turn a page into trustworthy fields, and those are different purchases.
Google's figure is real and we are citing it: each document page is equivalent to 258 tokens. It applies to Gemini 2.5. Google's media resolution guide gives Gemini 3 models a different table, 560 tokens for a PDF page at the default setting, 280 at low and 1,120 at high, and it says quality for PDFs "typically saturates at medium", which is the 560 setting. Either way, what that number describes is the page going in. It says nothing about the text coming back out, and the whole point of OCR is the text coming back out.
A page compresses on the way in and expands on the way out. Google charges you 258 tokens for a page whether it holds one sentence or a thousand words. The model then writes that content back to you one token at a time, at four to eight times the input rate. A typical dense business page emits somewhere near 750 output tokens, and a heavy table can pass 2,000.
Run the numbers on Gemini 2.5 Flash-Lite. Input for 1,000 pages: 258,000 tokens at $0.10 per million, about $0.03. Output for the same 1,000 pages: 750,000 tokens at $0.40 per million, about $0.30. Input is under 8% of the total. On Gemini 3.8 Flash, with 560 input tokens a page, input is still only 13%. The estimate everyone repeats is describing a rounding error.
The 750-output-tokens-per-page figure in our table is our stated assumption, not a Google number. Your pages decide your bill. Gemini 3.1 Pro is a preview model, and its $2.00 and $12.00 rates apply to prompts of 200,000 tokens or fewer, which every single-document OCR call will be. Output prices on every Gemini model include thinking tokens, so a Flash model with thinking left on can emit far more than the text it returns. Rates change, so confirm them on Google's current pricing page.
Token rates are Google's published figures, cross-checked against both the Gemini API pricing page and the Vertex AI pricing page. The per-1,000-pages columns are our arithmetic: 258 input tokens per page on Gemini 2.5 and 560 on Gemini 3 (Google's default PDF resolution), and an assumed 750 output tokens for a dense page. Swap in your own output figure and the last column moves with it.
| Model | Input tokens per page | Input, per 1M tokens | Output, per 1M tokens | Input cost, 1,000 pages | Output cost, 1,000 pages | Total, 1,000 pages |
|---|---|---|---|---|---|---|
| Gemini 2.5 Flash-Lite | 258 | $0.10 | $0.40 | $0.03 | $0.30 | About $0.33 |
| Gemini 3.1 Flash-Lite | 560 | $0.25 | $1.50 | $0.14 | $1.13 | About $1.27 |
| Gemini 2.5 Flash | 258 | $0.30 | $2.50 | $0.08 | $1.88 | About $1.95 |
| Gemini 3.5 Flash-Lite | 560 | $0.30 | $2.50 | $0.17 | $1.88 | About $2.04 |
| Gemini 3.8 Flash (same rate: 3.7, 3.6 Flash) | 560 | $0.75 | $3.75 | $0.42 | $2.81 | About $3.23 |
| Gemini 3.5 Flash | 560 | $1.50 | $9.00 | $0.84 | $6.75 | About $7.59 |
| Gemini 3.1 Pro Preview | 560 | $2.00 | $12.00 | $1.12 | $9.00 | About $10.12 |
Two things fall straight out of that table. The cheap models are absurdly cheap, and the Pro models are the wrong tool: Gemini 3.1 Pro costs roughly 31 times Flash-Lite to perform straight transcription, a task that needs no reasoning at all. If you are running Pro over scanned pages to save yourself a model-selection decision, that decision is costing you real money at volume. There is also a quiet discount worth knowing about. On the Gemini 3 models, Google states that natively embedded text pulled from a PDF is not charged as tokens, so a digital PDF with a genuine text layer costs only its page image. Scans and photographs have no text layer, so they tokenize as images and you pay in full. And if you send a scan as a JPEG rather than inside a PDF, Gemini 3 bills it at the image default of 1,120 tokens, twice the PDF figure, so wrap scans in a PDF or set media_resolution to medium. The newest Flash models are also the cheaper ones: Gemini 3.8, 3.7 and 3.6 Flash all bill $0.75 in and $3.75 out, half of what Gemini 3.5 Flash charges, which makes 3.5 Flash the one Gemini 3 model not to pick for OCR.
Published rates, taken from each vendor's own pricing page. AWS figures are US West (Oregon). Gemini figures are computed as described above. Read across a row, not down a column.
| What you need | Gemini | Azure AI Document Intelligence | AWS Textract | Google Document AI | Mistral |
|---|---|---|---|---|---|
| Plain OCR, text only | About $0.33 per 1,000 (2.5 Flash-Lite) | About $1.50 per 1,000 (Read) | About $1.50 per 1,000 | About $1.50 per 1,000 | About $4 per 1,000 |
| Structured field extraction | Prompt it and hope | About $30 per 1,000 (custom) | About $50 to $70 per 1,000 | About $30 per 1,000 (Form Parser) | About $5 per 1,000 |
| Confidence scores | Not returned | Yes | Yes | Yes | Opt-in: page, block or word |
| Bounding-box geometry | Not guaranteed | Yes | Yes | Yes | Yes |
| Billing unit | Tokens in and out | Pages | Pages | Pages | Pages |
| Cost varies with page density | Yes, output tokens scale | No | No | No | No |
| Deterministic output | No, generative | Yes | Yes | Yes | No, generative |
The top row and the middle rows point in opposite directions, and that tension is the whole decision. Gemini wins the price of reading a page outright. It loses every row that describes what you do with the result: no per-field confidence value to threshold on, no guaranteed geometry to draw a review box around, a bill that moves with how much text happens to be on the page, and a generative engine that can return two different answers to the same document. For the full cross-vendor picture, including the fees the headline rate hides, see our OCR API pricing comparison. If it is specifically Google's three products you are untangling, our Google Cloud Vision pricing page covers the one that returns no form fields at any price.
Buyers conflate them constantly, and the wrong pick can cost an order of magnitude in either direction. Here is the honest split.
Billed per token
A general multimodal model that reads documents very well as a side effect. Cheapest path from a page to its text, from about $0.33 per 1,000 pages. Returns generated text, no confidence scores, no contractual geometry. Best when a human or a downstream check will catch mistakes.
Billed per page
The purpose-built document platform. Enterprise Document OCR at about $1.50 per 1,000 pages, Form Parser and Custom Extractor at about $30. Returns confidence scores and geometry. Watch the roughly $438 a year idle hosting fee per deployed custom processor version.
Billed per unit
A classic OCR API at about $1.50 per 1,000 units, where a unit is one feature on one image. Returns text and nothing resembling a form field, at any price. If you need key-value pairs, this is the wrong Google product and no budget fixes that.
A realistic US mid-market workload: 35,000 pages a month of mixed business documents where you need named fields out of each one, not a wall of text.
$12
per month
About $0.33 per 1,000 pages. Returns text you then have to parse into fields, check, and correct yourself.
$68
per month
About $1.95 per 1,000 pages. Better reasoning over messy layouts, still no confidence scores.
$1,050
per month
About $30 per 1,000 pages, plus roughly $438 a year per deployed custom processor version, which accrues while idle.
$499
per month
The published 35,000 page plan, about $14 per 1,000 pages, with classification, validation, review, and export included.
Twelve dollars is a genuinely astonishing number and we are not going to talk you out of it. If you have engineers who will own a pipeline, and a downstream process that tolerates the occasional wrong value, Gemini Flash-Lite is the cheapest reader on the market by a wide margin. What the $12 does not buy is the classification step that decides what each page is, the parse from free text into named fields, the validation rule that catches a total which does not match the line items, the confidence signal that tells you which of those 35,000 pages a person should look at, the screen where they look at it, the audit trail your auditor will ask for, and the export that lands clean records in your ERP. That work is the difference between a model response and a finished document workflow, and it does not get cheaper because the tokens did. Price the whole job, not the API call.
Every vendor on one unit, including what the Gemini token rates actually work out to per page.
Azure, AWS Textract and Google Document AI per 1,000 pages, side by side.
$1.50 per 1,000 units, and why Vision returns no form fields.
$4 per 1,000 pages for OCR 4, and why the $1 figure is stale.
Every Claude model in dollars per 1,000 pages, from $5.31 to $64.64.
GPT-6 Luna at $0.62 per 1,000 pages, and the detail setting that can double a bill.
Every language model in one table, against the $1.50 dedicated OCR baseline.
The other token-billed reader: free weights, no API, and a real GPU bill.
Including the $438 a year idle processor hosting fee.
S0 rates for Read, Layout, prebuilt and custom extraction.
Per-feature rates from Detect Document Text through Analyze Lending.
$1.50 per 1,000 units for text, and why Vision never returns form fields.
Where a model beats OCR, where it hallucinates, and what each costs.
What a document extraction API returns when the pipeline is already built.
An honest roundup of the options for US teams.
A token estimate cannot tell you whether the fields come out right. Upload one of your real documents, look at what comes back, and then decide whether you are buying a model call or a finished pipeline.
