OpenAI does not sell OCR and does not publish a per-page price: there is no OCR endpoint and no invoice model, so you send a page to a vision-capable chat model and pay for tokens. On a normal US Letter invoice page that puts 1,000 pages between $0.47 on GPT-5 nano and $29.48 on GPT-5.5, with GPT-5 mini around $2.06. AWS Textract and Azure both charge a flat $10.00 per 1,000 pages, so most OpenAI models undercut them on raw inference. The surprise is which side of the request costs more: the JSON you ask back is 59 to 77 percent of the bill, not the image you send in.
Upload your invoices
Drop files here or click to upload
Up to 50 files
Uploading...
Every other vendor on this list quotes a price per 1,000 pages. OpenAI quotes a price per million tokens, and a page is not a fixed number of tokens. Four things make the conversion messy.
OpenAI has no OCR endpoint and no prebuilt invoice model. The vision-capable chat models read images, so OCR is something you prompt for rather than a service you buy. That means no per-page SKU exists to compare against Textract or Azure.
A page is converted to tokens based on its resolution. Scan the same invoice at 300 DPI instead of 150 and the image is four times the pixels, which changes the token count and the bill. Two teams can run the same volume and pay very different amounts.
GPT-4o and GPT-4.1 use the older tile method: 85 base tokens plus 170 per 512-pixel tile. The GPT-5 family uses 32 by 32 pixel patches against a 1,536-patch budget, then applies a per-tier multiplier. The same page costs different tokens on each.
Reading the page is cheap. Writing back structured JSON with line items is not, because output is billed at four to eight times the input rate. On GPT-5 the returned JSON is 77 percent of the cost of the request.
Below is the full derivation, using OpenAI published rates and OpenAI own published token formulas. Every assumption is stated so you can rerun it with your own page sizes.
GPT-5 mini lands at roughly $2.06 per 1,000 invoice pages at standard rates, or $1.03 through the Batch API. That undercuts Mistral OCR at $4 and both cloud invoice APIs at $10.
GPT-5.5 works out to about $29.48 per 1,000 pages against $0.47 for GPT-5 nano. Model choice matters more than any other cost decision you will make in this pipeline.
OpenAI Batch pricing is a flat 50 percent discount across all the models listed. Invoice processing is almost always asynchronous, so this discount is usually free money.
Because output is 59 to 77 percent of the bill, cutting fields you do not use out of your JSON schema saves more than downscaling the scan. Most teams optimize the wrong end.
Independent testing puts GPT-4o at 98 percent on header fields but only 57 percent on line items. If you need the table rows, that gap costs more than the tokens do.
No schema, no validation, no confidence scores, no review screen, no Excel export, no retries. Those are yours to build and maintain on top of the per-token rate.
The token rates are published. Turning them into a monthly number takes four inputs you control.
Open a representative scan and note the width and height in pixels. A US Letter page at 150 DPI is 1,275 by 1,650. At 300 DPI it is 2,550 by 3,300, which pushes further past the patch budget and gets downscaled anyway.
Tip: Scanning above 200 DPI usually buys you nothing here, because the model downscales anything over the 1,536-patch budget before it reads it.
For GPT-4o and GPT-4.1, count 512-pixel tiles after the resize and use 85 plus 170 per tile. For the GPT-5 family, count 32 by 32 patches, cap at 1,536, then multiply by 1.0 for the flagship, 1.62 for mini, or 2.46 for nano.
Write out the JSON you actually want back and count it. A header-only extraction is around 150 tokens. Header plus ten line items is around 700. This is the number that decides your bill, so do not guess it.
Price the developer days for schema design, validation, retries, and export, plus the ongoing time to fix layouts that break. Put that total beside a flat plan that already includes all of it.
If results within 24 hours is acceptable, route everything through the Batch API and halve the inference line. Very few invoice workflows genuinely need synchronous responses.
Token pricing is excellent value if you have engineers and a reason to build. It is poor value if what you actually wanted was invoices in a spreadsheet.
If your stack already calls the OpenAI API, adding document extraction costs you no new vendor, no new contract, and no new security review. That convenience is worth real money.
A general model handles a purchase order, a delivery note, and a handwritten receipt with one prompt change. A prebuilt invoice model cannot, and a custom model needs training data.
At tens of thousands of pages where you only need vendor, date, number, and total, GPT-5 nano through Batch is about $0.24 per 1,000 pages. Nothing purpose-built comes close on price.
If you need the table rows and you do not have developers, token pricing is a false economy. The 57 percent line-item accuracy and the missing review workflow will cost you more than the API ever saves.
These are the published standard-tier rates for every vision-capable GPT model, taken from the OpenAI API pricing documentation on 10 August 2026. Prices are per 1 million tokens. Batch pricing is a flat 50 percent discount on all of them.
| Model | Input / 1M | Cached input / 1M | Output / 1M |
|---|---|---|---|
| GPT-5.5 (under 272K context) | $5.00 | $0.50 | $30.00 |
| GPT-5 | $1.25 | $0.125 | $10.00 |
| GPT-5 mini | $0.25 | $0.025 | $2.00 |
| GPT-5 nano | $0.05 | $0.005 | $0.40 |
| GPT-4.1 | $2.00 | $0.50 | $8.00 |
| GPT-4.1 mini | $0.40 | $0.10 | $1.60 |
| GPT-4.1 nano | $0.10 | $0.025 | $0.40 |
| GPT-4o | $2.50 | $1.25 | $10.00 |
| GPT-4o mini | $0.15 | $0.075 | $0.60 |
Note the shape of these numbers. Output is billed at four to eight times input on every single model. That one fact drives everything else on this page.
To turn tokens into a page price you have to state your assumptions, so here are ours in full. One page per request. A US Letter page scanned at 150 DPI, so 1,275 by 1,650 pixels. Detail set to high. A 200-token instruction prompt. And 700 output tokens, which is about what a header plus ten line items comes to in JSON. Standard tier, no prompt caching.
Image tokens work out to 765 for the tile-based models and 1,496 patches for the GPT-5 family, because a full page exceeds the 1,536-patch budget and gets downscaled to 1,088 by 1,408 first. The mini and nano tiers then multiply that patch count by 1.62 and 2.46 respectively.
| Model | Image tokens | Input cost | Output cost | Total / 1,000 pages | With Batch | Output share |
|---|---|---|---|---|---|---|
| GPT-5 nano | 3,680 | $0.19 | $0.28 | $0.47 | $0.24 | 59% |
| GPT-5 mini | 2,424 | $0.66 | $1.40 | $2.06 | $1.03 | 68% |
| GPT-4.1 | 765 | $1.93 | $5.60 | $7.53 | $3.76 | 74% |
| GPT-5 | 1,496 | $2.12 | $7.00 | $9.12 | $4.56 | 77% |
| GPT-4o | 765 | $2.41 | $7.00 | $9.41 | $4.71 | 74% |
| GPT-5.5 | 1,496 | $8.48 | $21.00 | $29.48 | $14.74 | 71% |
Two honest caveats. First, OpenAI publishes image multipliers only for the mini and nano tiers, so we treat the flagship models as one token per patch, which matches the documented example of a 1,024 by 1,024 image costing 1,024 tokens. Second, we have deliberately left GPT-4o mini, GPT-4.1 mini and GPT-4.1 nano out of this table: their image token multipliers are not published in the vision guide, and we would rather show a gap than a number we cannot source.
Here is where the token math lands against vendors that do quote a page rate. All competitor figures are the ones verified on our individual pricing pages.
| Service | Cost / 1,000 invoice pages | Returns invoice fields? | Line items? | Field accuracy |
|---|---|---|---|---|
| GPT-5 nano | $0.47 | Only if you prompt and validate | Unreliable | Not independently tested |
| GPT-5 mini | $2.06 | Only if you prompt and validate | Unreliable | Not independently tested |
| Mistral OCR 4 | $4.00 | No, returns markdown | No | Not in this benchmark |
| GPT-4.1 | $7.53 | Only if you prompt and validate | Weak | Not independently tested |
| GPT-4o | $9.41 | Only if you prompt and validate | 57% to 63% | 90.5% to 98% |
| AWS Textract Analyze Expense | $10.00 | Yes, 40+ normalized fields | Yes, 82% | 78% |
| Azure prebuilt-invoice | $10.00 | Yes, native | Yes, 87% | 93% |
The pattern is clear once you line them up. OpenAI wins on price and loses on structure. Textract and Azure cost more per page but hand back a typed invoice object with line items already grouped, which is a large amount of code you do not write.
The best independent data available is a Businessware Technologies benchmark published in January 2025. It scored five tools against human-verified ground truth on scanned invoices in JPG, PNG and PDF without text layers, spanning at least three layout types and documents issued between 1971 and 2020.
| Tool | Field accuracy | Line-item extraction |
|---|---|---|
| GPT-4o with third-party OCR | 98.0% | 57.0% |
| Azure Document Intelligence | 93.0% | 87.0% |
| GPT-4o with image input | 90.5% | 63.0% |
| Google Document AI | 82.0% | 40.0% |
| AWS Textract | 78.0% | 82.0% |
Read those two columns together, because separately they mislead. GPT-4o with an OCR pre-pass is the most accurate tool tested on header fields, beating Azure by five points and Textract by twenty. On line items it is second from last, at 57 percent, roughly thirty points behind Azure. A general model reads text superbly and reasons about table structure poorly, which is exactly what you would expect.
The caveats matter. This is January 2025, and every vendor has shipped updates since, OpenAI most of all. The dataset is deliberately hard, low-quality scans with no text layer, so absolute numbers rise on clean digital PDFs. The benchmark authors also note that no tested solution was production-ready without additional configuration, which is the honest summary of this whole category.
The token cost is the smallest line in the budget. A working invoice pipeline on the OpenAI API also needs a JSON schema you design and version, prompt engineering per document family, retry logic for malformed output, numeric validation so line items sum to the total, confidence scoring so a human knows what to check, a review interface, storage, and an export path into your accounting system. Teams routinely spend six to ten developer weeks on this before it survives real vendor mail.
At 2,000 pages a month, GPT-5 mini costs about $4. Two developer weeks at a realistic US loaded rate costs several thousand dollars. The API price is not the decision.
You can send a PDF directly through the Responses API as base64, a Files API ID, or a URL. On vision-capable models the API extracts both the text layer and a rendered image of each page, and puts both in context. That is convenient and it doubles your token exposure, because you pay for the extracted text and the page image.
Each file must be under 50 MB, and the combined limit across all files in one request is also 50 MB. The detail parameter accepts auto, low or high, and only applies to the Responses API, not Chat Completions. Auto resolves to high on GPT-5.6 and later and to low on earlier models, which is a meaningful billing difference to be aware of if you never set it explicitly. Text extracted from the PDF is included regardless of the detail setting.
One practical trap: a 40-page PDF sent in a single request puts 40 page images in context at once. That is legal but expensive and usually less accurate than 40 single-page requests, because the model has more to attend to. Split documents page by page.
No. OpenAI has never shipped a dedicated OCR endpoint or a prebuilt invoice model. What exists is vision capability inside the general chat models, reached through the Responses API or Chat Completions with an image or file input. There is no equivalent of Textract Analyze Expense or the Azure prebuilt-invoice model in the OpenAI catalog, which is why no per-page price exists to quote.
For invoice extraction specifically, GPT-5 mini is the sensible default: about $2.06 per 1,000 pages, strong text reading, and a 68 percent output share you can trim. Move up to GPT-5 or GPT-4.1 when documents are dense or multi-column and the mini tier starts missing fields. GPT-5 nano is worth testing only for header-only extraction at high volume. GPT-5.5 is difficult to justify for OCR, since reading a page rarely needs frontier reasoning and you pay 14x the mini rate for it.
The underlying models are the same, but the economics and the controls are not. ChatGPT is a flat subscription with no per-page metering, which makes it fine for a handful of documents and unusable as a pipeline: no schema enforcement, no batch, no audit trail, and no way to guarantee the output format. The API gives you structured outputs, the Batch discount, and reproducibility. If you are processing invoices as a business process rather than a one-off, the API is the only sensible path, and a purpose-built tool is often better than either.
GPT-5 is cheaper on input, $1.25 against $2.50 per million, and identical on output at $10.00. But GPT-5 uses the patch method and burns 1,496 image tokens on a page where GPT-4o uses 765, so on our test page the two land within thirty cents of each other: $9.12 against $9.41 per 1,000 pages. The real saving is not GPT-5 over GPT-4o, it is GPT-5 mini over both, at roughly a quarter of the cost.
It depends entirely on whether you have developers. If you do, GPT-5 mini through the Batch API at about $1.03 per 1,000 pages is the lowest software cost available, and the build is the point. If you do not, a flat plan that already includes field extraction, line-item capture, validation, and Excel, CSV, JSON and QBO export removes the six to ten developer weeks that dominate the real total. Our own plans start at $49 a month for 2,500 pages, which is about $19.60 per 1,000 pages all-in, with nothing left to build.
Last updated August 2026. OpenAI token rates verified against the official OpenAI API pricing documentation on 10 August 2026. Accuracy figures attributed to Businessware Technologies, January 2025.
There is no official per-page rate, because OpenAI bills tokens rather than pages. On a US Letter page at 150 DPI returning header fields plus ten line items, 1,000 pages costs about $0.47 on GPT-5 nano, $2.06 on GPT-5 mini, $7.53 on GPT-4.1, $9.12 on GPT-5, $9.41 on GPT-4o and $29.48 on GPT-5.5. The Batch API halves every one of those.
No. OpenAI has no dedicated OCR endpoint and no prebuilt invoice model. OCR is something you prompt a vision-capable chat model to do through the Responses API or Chat Completions. That is why there is no per-page SKU to compare against AWS Textract Analyze Expense or the Azure prebuilt-invoice model.
GPT-5 mini is the sensible default for invoice work at about $2.06 per 1,000 pages. Step up to GPT-5 or GPT-4.1 for dense multi-column layouts where the mini tier drops fields. GPT-5 nano suits header-only extraction at volume. GPT-5.5 is hard to justify, since page reading rarely needs frontier reasoning at 14 times the mini rate.
For header fields, yes. Businessware Technologies measured GPT-4o with an OCR pre-pass at 98 percent field accuracy in January 2025, the highest of the five tools tested. For line items it scored 57 percent, well behind Azure Document Intelligence at 87 percent and AWS Textract at 82 percent. If you need the table rows, that gap is the problem, not the price.
OpenAI bills output at four to eight times the input rate on every model. A page image is 765 to 3,680 tokens, but the JSON you ask back with line items is around 700 tokens billed at the higher rate. Across the models in our table, output is 59 to 77 percent of the total. Trimming unused fields from your schema saves more than downscaling the scan.
Yes. Batch pricing is a flat 50 percent discount relative to standard rates across all the GPT models listed, and it applies to image and file inputs. Since invoice processing rarely needs a synchronous answer, routing work through Batch is usually a straight halving of your inference bill for the cost of waiting up to 24 hours.
AWS Textract Analyze Expense and the Azure prebuilt-invoice model both charge a flat $10.00 per 1,000 invoice pages. Every OpenAI model except GPT-5.5 comes in cheaper, and GPT-5 mini is roughly a fifth of the price. The tradeoff is structure: the cloud APIs return a typed invoice object with grouped line items, while OpenAI returns whatever your prompt and validation code can enforce.
Yes, through the Responses API as base64, a Files API ID, or a URL. On vision-capable models the API puts both the extracted text layer and a rendered image of each page in context, so you pay for both. Each file must be under 50 MB and the combined limit per request is also 50 MB. Split multi-page documents into single-page requests: it is usually cheaper and more accurate.
Everything that makes it a product. You design and version the JSON schema, write the prompts, handle malformed output, validate that line items sum to the total, add confidence scoring, build a review screen for low-confidence fields, store the results, and build the export into your accounting system. Six to ten developer weeks is a realistic estimate before it survives real vendor mail.
On inference alone, yes. On total cost, usually not unless you already have developers and a reason to build. GPT-5 mini at 2,000 pages a month costs about $4 in tokens. A flat plan at $49 a month, roughly $19.60 per 1,000 pages, already includes extraction, line items, validation, review and Excel, CSV, JSON and QBO export. The developer weeks are what decide it.
The flat $10 per 1,000 pages alternative, with every AWS rate verified.
Azure prebuilt-invoice cost per page, and the 93% accuracy behind it.
The other token-metered option: Gemini per-page math from $0.27 per 1,000.
Flat $4 per 1,000 pages, or $2 with Batch, for markdown output.
Both clouds charge exactly $10 per 1,000 pages, so accuracy decides.
Prefer an API? Pull invoice data to JSON without per-token metering.
Capture full line-item tables, the exact thing GPT-4o scores 57% on.
Get typed invoice JSON without writing a schema or prompt yourself.
Start turning your invoices into clean, structured spreadsheet data.
USD
per month
billed as
$288 yearly
Choose speed vs accuracy when extracting
| Base AI Faster | 2,500 pages |
| Pro AI Best accuracy | 500 pages |
Scale invoice extraction across your whole team with automation.
USD
per month
billed as
$888 yearly
Choose speed vs accuracy when extracting
| Base AI Faster | 10,000 pages |
| Pro AI Best accuracy | 2,000 pages |
Enterprise‑grade invoice extraction, security, and controls.
USD
per month
billed as
$ yearly
Choose speed vs accuracy when extracting
| Base AI Faster | pages |
| Pro AI Best accuracy | pages |
