Textract vs Signature Detector: cost comparison

This is a price and output-shape comparison. It is not an accuracy benchmark. Signature Detector does not publish a score against Amazon Textract, Azure Document Intelligence, or Google Document AI.

The figures below are list prices from the vendor pages linked at the end, checked while writing this post. Regions, commitments, and free tiers change the bill. Recompute them before you buy anything.

The unit

Signature Detector charges 1 credit per image and 1 credit per PDF page when the file is not served from cache. A credit is $0.0001 ($10 for 100,000). Images use imageUrl or imageBase64. PDFs use pdfUrl or pdfBase64 on the same POST https://signature-detector.com/api/detect. A PDF's pages, count, and creditsUsed are top-level fields. creditsUsed equals the page count.

Textract, Azure, and Google also charge per page. A 100,000-page corpus is $10 on Signature Detector whether you send page images or PDFs, as long as none of those files are cached.

Repeated requests for the same image bytes, or the same PDF bytes, return cached: true and do not spend a credit. A cached PDF sets creditsUsed to 0. A failed detection refunds the credit. New accounts start with 10 credits. The smallest purchase is $1 (10,000 credits). Credits do not expire. A PDF whose page count exceeds the balance returns HTTP 402 and is not charged.

100,000 pages

Amazon's Textract pricing example 8 detects signatures on 100,000 pages with Analyze Document Signatures. The worked total on that page is $350. Dividing $350 by 100,000 gives $0.0035 per page. One line in that example writes $0.00035. The $350 total and example 9 (below) both fit $0.0035, so this post uses $0.0035. The examples are for US West (Oregon).

$10 versus $350 is a 35 times difference for this unit. It is not a 35 times difference for a Textract call that also turns on Forms or Tables. Those features are priced separately and the examples on Amazon's page add them together.

Larger volumes

Textract Signatures drop to $0.0014 per page after the first 1 million pages in a month. Example 9 on the pricing page processes 5 million pages: $0.0035 × 1,000,000 = $3,500, plus $0.0014 × 4,000,000 = $5,600, total $9,100.

Google's Document AI pricing lists custom extractor at $30 per 1,000 pages up to 1,000,000 pages, then $20 per 1,000. Their Form Parser example, which uses the same tier shape, bills the first 1,000,000 at the higher rate and only the pages above that at the lower rate. For 5,000,000 pages that is $30,000 + $80,000 = $110,000. Committed-use columns on that page are lower. Hosting a custom processor version is listed at $0.05 per hour ($438 for a version left deployed for a year) and is not inside the per-page number.

Azure's East US retail meter S0 Custom Pages is $30 per 1,000 pages, effective 1 June 2024. S0 Custom Generative Pages and S0 Batch Custom Extraction Pages are also $30 per 1,000 on that list. There is no automatic volume discount on the S0 custom meter, so 5,000,000 pages at list price is $150,000. Commitment tiers are cheaper if you prepay. The East US "Commitment Tier Custom Azure 100K" unit is $2,400 per month, which is $24 per 1,000, with overage at $24 per 1,000. Microsoft's 1 June 2024 post is the announcement that custom extraction moved from $50 to $30 per 1,000 pages.

At 1,000,000 uncached pages, whether each page is an image or a page inside a PDF, Signature Detector is $100. Textract Signatures is $3,500. Azure S0 custom extraction is $30,000. Google custom extractor is $30,000.

At 5,000,000 uncached pages, Signature Detector is $500. The other totals are the ones above. Cached repeats are not in these totals.

What you are buying

Textract Signatures detects handwritten signatures, electronic signatures, and initials, and the pricing page says OCR text is part of that feature. You pass FeatureTypes: ["SIGNATURES"] to AnalyzeDocument. A SIGNATURE block includes location and confidence. Textract geometry is a bounding box plus a polygon, and the box values are ratios of the page width and height, not pixels. Block confidence is documented from 0 to 100. New AWS customers get 1,000 signature-only pages per month free for three months.

Google custom extractor can add a signature entity. The signature detection doc shows a found signature as mention text Detected with signature_value: true. If nothing is found, the entity is left out of the response rather than returned as false. Because the field is derived, textAnchor and pageAnchor are not populated, so that entity does not include a box. You label a schema, and the doc lists specific foundation-model versions. The derived-fields section is marked preview.

Azure custom extraction is a model you train to emit the fields you labeled. The $30 price is for that custom meter, not for every Azure document model. Prebuilt models are listed separately on the pricing page. The FAQ there says custom neural training is free for the first 10 hours, then $3 per hour.

Signature Detector returns pixel boxes (x1, y1, x2, y2), a confidence score, class, and class_name. The sample in the docs uses confidence 0.192. Image requests accept an optional threshold from 0 to 1 and default to 0.1. PDF detection always uses 0.1. There is no OCR text, no polygon, and no training step. Do not compare 0.192 with a Textract confidence of 19.2 or 90 as if they shared a scale.

When the cheaper call is the wrong call

Use Textract, Document AI, or Document Intelligence when the pipeline also needs words, key-value pairs, or tables from the same page. Paying $0.0035 and throwing away the text is wasteful only if you did not need the text. Paying $0.0001 and then running a second OCR product can erase the gap. This API is the right size when the question is where the signature marks are.

The same numbers are summarized on Textract signature alternative. Request fields are on Signature detection API.

Sources: Amazon Textract pricing, Textract Block, AnalyzeDocument, Document AI pricing, Google signature detection, Azure Document Intelligence pricing.