Key takeaways

  • AWS announced Textract Insights general availability on November 20, 2026, with launch availability in us-east-1 and us-west-2 according to its pricing page.
  • The asynchronous StartDocumentInsights workflow can split a mixed packet of up to 3,000 pages and return blueprint fields with confidence scores and source page coordinates.
  • AWS’s examples price Extraction at $8.50 and Extraction + Page summaries at $9.50 per 1,000 pages. Custom output includes 30 fields at $40 per 1,000 pages, with add-ons beyond that.
  • Insights supports six languages, not the 40 announced for Textract Core. Source coordinates and confidence scores still need validation on the buyer’s documents.
01

What AWS announced and where it is available

AWS announced the general availability of Amazon Textract Insights on November 20, 2026: a generative AI feature accessed through a new asynchronous Textract API. Builders describe the data they want, submit a document or mixed packet from Amazon S3, and receive structured values with confidence scores and source grounding. AWS’s pricing page lists availability in US East (N. Virginia), us-east-1, and US West (Oregon), us-west-2, with the same prices in both Regions.

The announcement names StartDocumentInsights for submission, CreateProject for projects and CreateBlueprint for extraction blueprints. It does not identify an underlying foundation model or a model API ID. This is a managed document-processing release, not downloadable model weights or a general chat-model endpoint. AccessAllGPT has not checked account-level access.

02

One packet, multiple blueprints

A blueprint is a JSON Schema with a natural-language description for each requested field. Its inference type specifies whether to read a value verbatim or infer it from the document. Builders can start from public AWS blueprints or create their own, then place the expected document types in a project.

With splitting enabled, AWS says one file of up to 3,000 pages can be separated into logical sub-documents, matched to blueprints and processed into results in the specified S3 bucket. The launch’s mortgage-packet example combines forms, payslips, bank statements and handwritten notes. It illustrates the workflow; it is not an AccessAllGPT trial.

Both explicit and inferred fields receive confidence scores and grounding that identifies a page and bounding-box coordinates. That makes a result easier to inspect than an unsupported answer from a generic prompt. It does not establish that the extraction is correct, that every inference follows from the cited region, or that a particular confidence threshold is calibrated for your files.

03

Pricing depends on the output, not just the page count

AWS’s Insights pricing examples list Extraction at $8.50 per 1,000 pages and Extraction + Page summaries at $9.50 per 1,000 pages. Custom output, including 30 fields, is $40 per 1,000 pages. These are different output types: the headline’s starting rate is not the price for every custom blueprint job. AWS states there is no minimum commitment.

For custom fields beyond the included 30, the published example charges $0.50 per 1,000 page-fields for a Standard field with an instruction of up to 120 characters, or $1.25 for a Dense field with 121–600 characters. AWS says the lightest descriptions are included first and the densest extra fields are billed. Its worked example of 1,000 pages with 40 fields, including four extra Standard fields and six extra Dense fields, totals $49.50. Figure captions are a separate add-on, illustrated at $0.80 per 1,000 figures.

The page also illustrates a mixed job in which blueprint-matched pages use Custom pricing and unmatched pages use Extraction pricing. Our analysis: estimate spend using the expected output mix and field descriptions, not a single starting rate multiplied across every document. Treat S3, other supporting services and human review as separate budgeting questions; these examples are not an all-in project quote.

04

Six Insights languages are not 40 Core languages

Insights supports PDF, PNG, JPEG and TIFF inputs and processes documents in English, French, German, Italian, Portuguese and Spanish, according to the launch post. AWS separately announced that Textract Core’s DetectDocumentText and AnalyzeDocument APIs now cover 40 languages, with improved dense-table and handwriting handling. The 40-language claim does not describe Insights.

The same post says Core has expanded to Tokyo and Zurich, bringing its in-region processing coverage to 18 Regions, while explicitly noting that Insights is not yet available in those two new Regions. Core’s existing API behavior and regional footprint should not be used as a proxy for this new service’s access or routing.

05

Cross-Region inference changes the residency question

AWS says Insights automatically routes inference to available compute within the geography where the request originated; its example keeps United States requests within the United States. Core APIs, by contrast, continue processing in the Region called. Buyers who require a single-Region processing commitment should resolve that difference before uploading documents. We did not establish an exhaustive Insights destination list from the general endpoint reference linked by the announcement.

The launch describes separate encryption controls for projects, blueprints and analysis jobs. A project key does not automatically apply to its jobs. AWS says intermediate job data is transient and deleted when a job completes, while the output bucket’s default encryption applies to results. These are vendor statements about the described workflow, not an independent retention or compliance audit.

06

Why it matters and what to try next

For teams assembling OCR, packet splitting, classification, model prompts and review into one pipeline, the practical change is a managed API that combines those steps and returns inspectable field-level evidence. Whether it is a better purchase depends on extraction quality, review workload, processing time and total cost on the team’s own document distribution. The launch supplies no independent benchmark or numerical latency result establishing that trade-off.

Start with synthetic, non-sensitive invoices and supporting letters in deliberately mixed order. Define expected fields before submission, distinguish verbatim from inferred answers, and check packet boundaries, missing values and the actual bounding boxes against a manually checked reference. Tune review thresholds on a representative labeled sample rather than accepting a score as a guarantee. This is our proposed acceptance exercise, not testing we performed.

07

A first document-AI buying check

Proposed acceptance criteria for a trial; no AccessAllGPT API results are implied.

Entries stay in this browser tab and are not submitted to AccessAllGPT. Blank responses are copied as [Unresolved].

Choose Extraction, summaries or Custom explicitly. Count matched pages, included and extra fields, instruction lengths, optional captions and supporting-service charges.

Score packet splitting and field accuracy against checked synthetic references; inspect source coordinates and tune thresholds on representative labeled documents.

Confirm launch Region, six-language coverage, cross-Region inference destinations and separate job encryption settings before sending real files.

Primary sources

  1. Announcing Amazon Textract Insights: Verifiable generative AI for documentsAmazon Web Services · Reviewed: Publication date and timestamp; general availability; JSON Schema blueprints and StartDocumentInsights; 3,000-page limit; explicit and inferred fields; Core comparison; encryption and cross-Region inference; six-language and input-format list · Retrieved · Supports: First-party launch date, API workflow, vendor-described grounding and confidence scores, input limits and languages, Core distinctions and data-handling claims. Not independent extraction accuracy, confidence calibration, latency, compliance or account-level access.
  2. Amazon Textract pricing — Textract InsightsAmazon Web Services · Reviewed: Textract Insights tab; availability in Northern Virginia and Oregon; Extraction and summaries worked examples; Custom 30-field inclusion, Standard and Dense add-ons, figure captions and mixed-output example · Retrieved · Supports: Published launch Regions and vendor pricing examples for the named output types and add-ons. These are public rates and worked examples, not an AccessAllGPT invoice or an all-in deployment-cost estimate.
  3. Amazon Textract endpoints and quotasAWS General Reference · Reviewed: General service endpoint listing and quota reference reached through the launch’s documentation link; checked for Insights-specific cross-Region destination information · Retrieved · Supports: Shows why the general Textract endpoint footprint must not be treated as an Insights availability or destination map. The retrieved reference did not establish an exhaustive Insights-specific inference destination list.

Limitations

Public-source briefing based on AWS materials. No authenticated AWS access, API call, document upload, hands-on extraction, invoice measurement, customer interview, independent accuracy benchmark or confidence-calibration study. No exhaustive Insights inference destination map or account-specific quotas were established. Rates are vendor examples and can change; AWS customer testimonials are not treated as independent verification.

Disclosures

Public-source reporting, not a sponsored test. AccessAllGPT Research is operated by NeuralArc and is independent of AWS. AWS did not review or endorse this briefing; publication-wide relationships are described on the disclosures page.

Further AccessAllGPT guidance

  1. Automation workflows and human approval boundaries
  2. AI API data retention and residency procurement gates
  3. What AI models cost
  4. Automation
  5. How this publication is made