Amazon Rekognition OCR Pricing vs Textract: What DetectText Costs and the 100-Word Limit

Amazon Rekognition DetectText costs $1.00 per 1,000 images and Amazon Textract DetectDocumentText costs $1.50 per 1,000 pages, so Rekognition looks like a 33% saving on OCR. It is not, because Rekognition will not read a document. AWS states in its own quota documentation that "DetectText can detect up to 100 words in an image" and that "Amazon Rekognition supports the PNG and JPEG image formats". A single-page invoice runs well past 100 words and usually arrives as a PDF, which Rekognition does not accept at all. The two services are priced close together because they do different jobs, not because they are substitutes.

This comes up constantly on teams costing a document pipeline on AWS, because Rekognition is the service most engineers meet first, its text detection genuinely works, and the price is right there in the console. The failure mode is not a bad result. It is a good result on the first three lines of the page and nothing after that, discovered two weeks into a build. Here are both rate cards, read from Amazon's own price list feeds on September 9, 2026, and the exact point where the cheaper service stops being an option.

What Amazon Rekognition DetectText actually costs

DetectText is a Group 1 image API, so it bills against the Rekognition image tier rather than anything document-specific. From Amazon's Rekognition price list feed, offer file 20260831092223 published August 31, 2026, US East (N. Virginia) on-demand:

Four tiers is generous by AWS standards, and the floor of $0.40 per 1,000 is lower than anything Textract offers. That floor is why the comparison keeps getting made. It is also why it is worth being precise about what you get for it.

What Amazon Textract costs for the same job

Textract's cheapest operation, DetectDocumentText, is $0.0015 per page for the first million pages a month and $0.0006 above that, from the Textract price list feed, offer file 20260831092230. That is $1.50 and $0.60 per 1,000 pages. Every other Textract operation costs more, sometimes a great deal more: TABLES is $15.00 per 1,000, FORMS is $50.00, and asking for FORMS with Custom Queries and TABLES in a single call is $80.00. The full sixteen-row matrix sits on our AWS OCR pricing page.

The 100-word cap is the whole decision

Everything else on that table is a preference. The word cap is a wall. AWS lists it under set quotas, the category the documentation describes as limits that "cannot be changed", alongside the maximum image size and the minimum face dimensions. It is not a default you can raise with a support case.

To put 100 words in context: a plain US commercial invoice with a header block, a bill-to address, eight line items and a totals section is usually somewhere between 150 and 400 words. A single page of a bank statement with thirty transactions is well over 300. A W-2 is over 200 before you count the boxes. Even a short delivery note tends to clear 100 once the addresses are in. There is essentially no business document category where you can rely on staying under the cap, and the failure is silent: you get a valid response containing the first 100 words, not an error telling you the rest was dropped.

The PDF restriction closes the door from the other side. Rekognition takes PNG and JPEG. If your documents arrive as PDFs, and in US accounts payable and lending they nearly always do, you would be rasterizing every page to an image before you could call the API at all, which adds a rendering step, a storage cost and a quality decision to a pipeline you chose for its simplicity.

So what is Rekognition DetectText for?

Text in photographs. Signs, street numbers, product labels, packaging, license plates, screenshots, the text baked into a marketing image. Short, scattered text in a scene, where there is no page structure to preserve and where 100 words is a generous ceiling rather than a tight one. It is genuinely good at that, and it is priced for the volume that use case generates, which is why the tiers run to 35 million images.

Content moderation pipelines use it constantly: scan an uploaded image for text, check the text against a policy, act. That workload sends millions of images and reads a handful of words from each, which is exactly the shape the pricing rewards. A document pipeline is the opposite shape, few inputs and a lot of words in each, and it gets the worst of both services if you point it at the wrong one.

What about Amazon Comprehend?

Comprehend comes up in the same conversation and belongs to a different stage. It never sees your image. It takes text that already exists and returns entities, key phrases, language, sentiment and PII, which is why AWS documents a pattern for feeding Textract output into Comprehend rather than choosing between them. If your source is a scan, Comprehend cannot start until an OCR service has finished, and you pay for both. Anyone comparing Comprehend against Textract is comparing two consecutive steps in the same pipeline.

Worth noting for the same reason: if the text you need is on a web page rather than inside a file, none of these services is the right shape at all. That is a scraping problem, and an API that returns clean, structured page data gets you there without an OCR bill.

The cost comparison that actually matters

Once you accept that Textract is the only one of the two that reads documents, the interesting comparison stops being Rekognition against Textract and becomes Textract against itself. Its price swings by 53x depending on which feature flags your code sends, and that swing dwarfs anything you could have saved by picking the cheaper service.

A pipeline that routes every page through FORMS because one document type needs key-value pairs pays $50.00 per 1,000 on all of them. Getting that routing right is worth more than any service-selection decision on this page. The related trap is splitting features across calls: Textract prices combinations below the sum of the parts, so asking for QUERIES and TABLES separately costs $30.00 per 1,000 pages where one call asking for both costs $20.00. We covered that and the other billing mechanics in why your Textract bill beats your estimate.

How to choose in about two minutes

Ask three questions in this order and you are done.

Is the input a page or a photograph? A page means a document with structure you care about: reading order, columns, tables, fields in fixed positions. A photograph means text incidental to a scene. Pages go to Textract. Photographs go to Rekognition.

Do you need values or characters? If what you want out is "invoice_total: 1420.00", no OCR service gives you that directly. Textract returns Block objects with types like WORD, LINE, CELL, MERGED_CELL and KEY_VALUE_SET connected by a relationships graph, and turning that into named fields is a parser you write, test against every layout you receive and maintain forever. That reconstruction layer is the largest hidden cost in most AWS document pipelines and it never shows up in a per-page comparison.

Does anything downstream depend on a confidence score? Both services return one, which is fortunate, because AWS's newer document service does not: we counted the word "confidence" in Amazon's Bedrock Data Automation document-output documentation on September 9, 2026 and it appears zero times. If your approval logic routes low-confidence fields to a human, that matters more than the rate. The arithmetic behind picking a threshold, including why a 98% per-field pass rate still sends a third of a twenty-field invoice batch to a person, is on our OCR confidence score page.

The short version

Rekognition DetectText is 33% cheaper than Textract at entry rate and cannot process a business document, because AWS caps it at 100 words per image and accepts only PNG and JPEG. Use it for text in photographs, where it is excellent and correctly priced. Use Textract for pages, and spend the attention you were going to spend on service selection on feature-flag routing instead, where the money actually is. If what you want at the end is typed fields with a confidence score on each one rather than a graph of Block objects to parse, that is a different product category again, and it is worth pricing before you write the parser.