Amazon Textract or Claude on Amazon Bedrock: which one should read the documents in your S3 bucket?

What each one returns

Textract returns Block objects. For a plain text detection call that means PAGE, LINE and WORD blocks, each with a bounding box and a confidence score. Turn on the analysis features and you also get KEY_VALUE_SET blocks for form fields, TABLE and CELL blocks for tables, QUERY and QUERY_RESULT blocks for the questions you asked, SIGNATURE blocks for where someone signed, and LAYOUT blocks that label paragraphs, titles, headers, footers and figures in reading order.

That output is built for pipelines. Every value has a location on the page and a score you can threshold, so a downstream system can accept "Invoice total: 4,820.00" at 99 percent and route it to a human at 70 percent. What Textract does not do is reason. Ask it "which of these line items are taxable" and you have to write that logic yourself.

Claude returns the opposite shape. The Bedrock Converse API takes the document and a question and gives back a text answer. With citations enabled, the answer arrives as citationsContent blocks that each name the page range the passage came from. You can ask for structured JSON in the prompt, but it is a generated answer, not a measured one, and there is no confidence score or bounding box.

• Textract: words, lines, form fields, table cells, query answers, signatures, layout, each with geometry and confidence.
• Claude: a prose answer, optionally with page citations into the document you sent.
• Textract reads scans; Claude reasons about content.

File limits that decide the architecture

Textract accepts JPEG, PNG, PDF and TIFF. The synchronous operations take a file of up to 10 MB and, for PDF and TIFF, one page only. Multi-page documents go through the asynchronous operations, which accept PDF and TIFF up to 500 MB and 3,000 pages, read the file from S3, and notify an SNS topic when the job finishes. PDFs cannot be password protected and XFA forms are not supported.

Because Textract is OCR, it does not care whether the PDF has a text layer. A photographed receipt, a faxed purchase order or a scanned signature page all work, in English, French, German, Italian, Portuguese and Spanish, with handwriting in English only. Text has to be at least 15 pixels tall, which is about 8 point at 150 DPI, and vertical text is not supported.

Claude through Converse accepts pdf, csv, doc, docx, xls, xlsx, html, txt and md, up to five documents per message and 4.5 MB each. That covers Office files Textract cannot read at all. But citations work on extracted text, so a scanned PDF with no text layer has nothing to cite, and anything over 4.5 MB has to be split or summarised in stages.

Ask questions about S3 documents with Claude on Amazon BedrockHow to preview files in Amazon S3 without downloading them

• Textract sync: 10 MB, one page for PDF and TIFF.
• Textract async: 500 MB and 3,000 pages, S3 input, SNS notification.
• Converse: 4.5 MB per document, five per message, Office formats included.
• Scans: Textract yes, Claude citations no.

Example: ask Textract a question about a multi-page PDF in S3

Textract Queries let you ask for a field in plain language rather than parsing key-value pairs yourself, up to 15 queries per page synchronously and 30 asynchronously. Query detection is English only. The script below starts an asynchronous analysis on a PDF in S3, polls for the result, and prints each query with its answer and confidence. Production code should subscribe to the SNS topic instead of polling.

cat > textract_queries.py <<'EOF'
import time
import boto3

textract = boto3.client("textract", region_name="us-east-1")

job = textract.start_document_analysis(
    DocumentLocation={"S3Object": {"Bucket": "amzn-s3-demo-bucket", "Name": "invoices/inv-10422.pdf"}},
    FeatureTypes=["QUERIES", "TABLES"],
    QueriesConfig={"Queries": [
        {"Text": "What is the invoice total?", "Alias": "total"},
        {"Text": "What is the payment due date?", "Alias": "due_date"},
    ]},
)

blocks, token = [], None
while True:
    kwargs = {"JobId": job["JobId"]}
    if token:
        kwargs["NextToken"] = token
    page = textract.get_document_analysis(**kwargs)
    if page["JobStatus"] == "IN_PROGRESS":
        time.sleep(5)
        continue
    if page["JobStatus"] != "SUCCEEDED":
        raise SystemExit(page["JobStatus"])
    blocks += page["Blocks"]
    token = page.get("NextToken")
    if not token:
        break

by_id = {b["Id"]: b for b in blocks}
for b in blocks:
    if b["BlockType"] == "QUERY":
        for rel in b.get("Relationships", []):
            for rid in rel["Ids"]:
                ans = by_id[rid]
                print(b["Query"]["Alias"], "=", ans["Text"], f"({ans['Confidence']:.0f}%)")
EOF
python3 textract_queries.py

What each one costs

Textract bills per page, and the price depends on which features you turn on. In US West (Oregon) for the first million pages a month, plain text detection is $0.0015 a page. Tables and Queries are $0.015 a page each, Forms is $0.05, Signatures $0.0035, and Layout is free when combined with Tables. Stack Forms, Tables and Queries on one call and the page costs $0.07. Invoices and receipts through AnalyzeExpense are $0.01 a page. Prices drop past a million pages, and new accounts get three months of 1,000 text pages and 100 analysis pages a month free. Check the pricing page before you budget; these figures are from October 2026.

Claude on Bedrock bills per token, in and out, and the document is most of the input. Our measurements for a 30-page text report came to about a cent per question on Claude Haiku 4.5 without prompt caching, and a tenth of that for follow-up questions with caching on. Amazon Nova Lite answered the same question for well under a tenth of a cent.

Put side by side, running Textract Queries over that 30-page report costs about $0.45 for up to 30 questions on every page, with coordinates and confidence, whether or not anyone reads the result. Asking Claude one question costs about a cent and returns one answer. Textract wins when you need every field from every page; a model wins when a person wants one answer now.

Amazon Bedrock pricing for document Q&AWhat a document chat costs on Amazon Bedrock, by model, with prompt caching

• Textract: per page, from $0.0015 for text to $0.07 for Forms, Tables and Queries together.
• Claude: per token, roughly a cent per question about a 30-page report on Haiku 4.5.
• Textract is paid whether or not anyone looks; a model is paid when someone asks.

When to use Textract

Choose Textract when the input is paper that has been scanned or photographed, when you need a value for every document in a batch, or when the result feeds a system that needs a confidence score and a location on the page. Accounts payable, claims intake, onboarding forms and lending packages are the classic cases, and Textract has dedicated AnalyzeExpense, AnalyzeID and Analyze Lending operations for them.

It is also the right tool when a human must be able to verify the extraction visually. A bounding box lets your review screen draw a rectangle around the field it read, which is a stronger audit trail than a sentence.

• Scanned or photographed documents with no text layer.
• Bulk extraction of the same fields from thousands of files.
• Downstream validation that needs confidence scores and coordinates.
• Invoices, receipts, US identity documents and mortgage packages.

When to use Claude on Bedrock

Choose a model when the question is not known in advance. "Does this agreement let us terminate for convenience" is not a field, it is a reading task. The same goes for summaries, comparisons between two versions, and anything where the answer depends on more than one place in the document.

Models also handle formats Textract cannot open. A Word contract, an Excel schedule or a Markdown runbook goes straight into Converse. And when the document already has a text layer, which is true of most PDFs generated by software, a model with citations gives you both the answer and the page it came from.

Amazon Bedrock Knowledge Bases on S3, or just ask the document?

• Ad hoc questions about one file, asked by a person.
• Summaries, comparisons and judgement calls across a document.
• Office formats and text-native PDFs.
• Answers that need a page reference rather than a coordinate.

Chaining them: Textract first, then the model

The common production pattern is both. Textract reads the scan into text and structure, that text goes to the model as a plain-text document, and the model answers questions or classifies the document. This fixes the scan problem for the model and the reasoning problem for Textract, and the Textract output stays available as the audit record.

If the files are text-native PDFs, skip the first step; sending the PDF straight to Converse is cheaper and keeps page citations. If you would rather not wire the pipeline yourself, Amazon Bedrock Data Automation packages extraction and generative output behind one API with S3 input and output, at the cost of another service to learn and price.

Automate S3 file workflows with Python and boto3

• Scans: Textract to text, then the model.
• Text-native PDFs and Office files: straight to the model.
• Keep the Textract JSON as the verifiable record.

Where Amazon Q and BucketDesk fit

Neither Textract nor a raw Converse call is something a finance or legal team uses directly. The products people compare at that level are index-based assistants and per-document tools. Amazon Q Business, which indexed S3 buckets behind a chat app, is closed to new customers, and AWS points new work to Amazon Quick, which also crawls and indexes the bucket before it can answer. Both answer across many files and neither extracts fields with coordinates the way Textract does.

BucketDesk Document AI is the per-document route. You open a file in a connected bucket, ask in plain language, and the answer comes back with a citation that highlights the supporting passage. The model runs on Bedrock in your AWS account, nothing is indexed ahead of time, and no file content is retained by BucketDesk after the session. It does not replace Textract for bulk extraction; it replaces the script a person would otherwise write to ask one contract one question.

Document AI on the features pageAmazon Bedrock data retention modes for business documents