Article

Textract Insights costs, file limits and when to use Core

For document-processing teams, compare Textract Insights page charges, packet limits and US processing rules before replacing Core.

Editorial illustration for Textract Insights costs, file limits and when to use Core: a document represents the research briefing. Not documentary evidence.

AWS announced Amazon Textract Insights’ general availability on November 20, 2026, giving teams handling invoice bundles, claims and onboarding files a new document-processing option. At current list rates, 10,000 pages cost US$85 for Extraction or US$400 for Custom with up to 30 fields, before extras. Keep Textract Core when you need synchronous results or individual table-cell evidence.

A blueprint is a list of fields and instructions describing what to extract. With splitting enabled and blueprints configured, Insights can separate mixed document packets and extract the requested fields. That makes the decision about the records your business needs—not simply the cheaper page rate.

What will 10,000 pages cost?

These published US-dollar rates apply in both supported US regions as of November 20, 2026.

Output selected

Per 1,000 pages

For 10,000 pages

Extraction

$8.50

$85

Extraction + Page summaries

$9.50

$95

Custom, up to 30 fields

$40

$400

These are alternative subtotals: pages ÷ 1,000 × rate. Assume one processing pass, no add-ons, credits, discounts or taxes. Storage, surrounding application services, development and human review are excluded. Do not add the Extraction base charge again to Custom pages.

For extra fields, AWS includes the 30 shortest field descriptions. Each remaining field adds $0.50 per 1,000 pages for instructions up to 120 characters, or $1.25 for 121–600 characters. Figure captions add $0.80 per 1,000 figures—not pages.

An illustrative mixed workload shows why page mix matters. Assume multiple distinct primary blueprints, no fallback, and 10,000 pages processed once: 5,000 receive Custom output with at most 30 fields; 5,000 unmatched pages receive Extraction only. Exclude summaries and captions.

Illustrative subtotal: (5,000 ÷ 1,000 × $40) + (5,000 ÷ 1,000 × $8.50) = $242.50. This uses AWS’s matched-versus-unmatched billing rule; the assumed match rate is not a measured result.

Crucially, an unmatched document can lack custom fields while the job still succeeds. A lower bill therefore does not, by itself, show better efficiency: reconcile the pages charged with the business records actually produced.

The listed Free Tier allowances cover Core APIs; they do not establish a free Insights allowance.

Which files fit the limits?

The fixed document limits depend on the features enabled. The headline 3,000-page allowance is not universal.

Check

Documented limit

File format

PDF, PNG, JPEG or TIFF; not DOC or DOCX

File size and overall job length

500 MB and 3,000 pages

Custom extraction without splitting

20 pages

Figure captioning enabled

1,500 pages

For custom extraction beyond 20 pages, AWS directs users to enable splitting. The splitter also requires blueprints: enabling it without them does not separate documents. Do not assume this guarantees that an arbitrary long, single document will divide into useful records.

Insights supports English, French, German, Italian, Portuguese and Spanish. Core’s broader text-detection language coverage does not carry over; Hindi, for example, is not on the Insights list.

Account quotas are separate and adjustable: defaults are 10 start requests and 10 status requests per second, plus 20 concurrent jobs in each of the extraction, generative and custom categories. An over-page-limit file can receive a job identifier and fail afterward. Submission accepted does not mean document processed.

Where will the documents be processed?

You can call Insights in N. Virginia (us-east-1) or Oregon (us-west-2). Input and output S3 storage buckets must be in that same request region. No India or EU request region is listed.

Storage location is not the whole processing boundary. AWS automatically routes inference—the AI processing—to N. Virginia, Ohio, N. California or Oregon from either supported endpoint. You do not select a routing profile.

For a team required to keep processing in one named region, same-region storage is therefore insufficient. AWS documents in-region processing for Core, while customer-managed encryption keys do not change Insights’ cross-region routing. Confirm acceptable processing locations before uploading records.

When should you keep Textract Core?

AWS’s Core-versus-Insights comparison identifies four reasons to retain the existing APIs:

  • A synchronous response. Core supports synchronous single-page calls; Insights document jobs run asynchronously.

  • Evidence for each table cell. Insights returns table values by row and column, but not a separate location and confidence score for every cell. Core’s AnalyzeDocument is the documented choice for highlighting individual cells in a review screen.

  • Every form field. Core’s Forms feature returns all key-value pairs; Insights custom extraction returns the fields listed in the blueprint.

  • Raw text alone. DetectDocumentText is the simpler Core option when you just need the words on the page.

Insights is worth evaluating when your task combines mixed document types, custom business fields and reusable output. But it is currently API-only, with results delivered asynchronously to S3, not a finished expense-approval application. Integration and review remain part of the project.

If you are comparing field-extraction services, the Clairic and Clairic-flash cost-and-limits guide covers a different billing model. Compare accepted records and correction work, not a page price against a token price as though they bought identical results.

If the real task is an assistant reading selected material rather than converting whole packets into fixed fields, see selecting PDF pages and spreadsheet cells with the Gemini API. That is a different workflow, not an interchangeable Insights interface.

What should you check before switching?

Start with representative documents you are permitted to process and known answers for the fields that matter. Check document boundaries, missing values, amounts and source locations before deciding which records may proceed without review.

Insights supplies field confidence scores and source locations to help with that review. A score is not a guarantee that a value is correct. Missing fields can be empty strings, so successful parsing is not evidence of a complete record.

Keep blueprint-match confidence separate from field confidence. With one blueprint and no fallback, matching is skipped and match confidence is always 1. That configuration cannot demonstrate perfect classification. A fallback result uses match confidence 0; it is not a statement that every extracted field is wrong.

Inspect each document’s match status, including unmatched and fallback results, rather than relying only on overall job success. Build job-status polling with backoff; AWS’s performance guidance describes variable latency, not a fixed completion-time guarantee.

The useful adoption test is whether the required fields reach your downstream system with acceptable correction work and processing locations. Budget for that complete path. Keep Core where its response mode or evidence detail is essential, and pilot Insights where managed splitting and custom fields solve a specific gap.