AWS says structured outputs for Anthropic Claude models on Amazon Bedrock are generally available through both bedrock-runtime and bedrock-mantle. Availability covers all Claude models where the respective model is available. The feature lets developers define a response structure with JSON Schema and enforce its supported rules during generation.

Structure is enforced as tokens are generated

Structured outputs uses constrained decoding to produce JSON that follows the schema, including field types, required fields, and permitted fields. AWS describes less need for custom parsing and retries caused by schema violations, alongside fewer output tokens. The company says this can potentially lower inference costs for high-volume workloads; the article does not establish independently measured savings.

The listed uses include extracting data from financial reports, compliance documents, clinical notes, invoices, contracts, emails, and transcripts. Other examples include enriching product catalogs, validating agent tool arguments, and classifying or routing tickets, leads, and content. AWS also describes Claude Haiku 5.5 as handling summaries, compactions, database queries, and classification requests.

APIs and tool inputs

Developers can use the feature through Boto3 and bedrock-runtime APIs, or through the Messages API and bedrock-mantle. It supports batch inference and streaming responses delivered as typed events. Cross-Region inference is supported on bedrock-runtime. With InvokeModel, the schema is supplied as a JSON object in output_config.format within an Anthropic-native request body, rather than as a JSON string.

On bedrock-mantle, strict: true constrains tool inputs to the specified schema, including field types, permitted values, and allowed fields. AWS describes Mantle as supporting agentic applications with server-side tools, background inference, and stateful conversations. Web search and code interpreter are named as server-side tool examples. An accompanying sample notebook provides additional API examples, including the Messages API.

Schema compilation and reuse

Bedrock validates schemas against a supported subset of JSON Schema Draft 2020-12. It compiles new schemas into grammar artifacts, which can increase first-request latency. Compiled grammars are cached within the account for 24 hours from first access. Structural changes trigger recompilation, while changes only to name or description fields preserve the cached grammar.

Every object must set additionalProperties: false for the schema to be accepted. AWS recommends descriptive property names and descriptions, with enum for fields that have a limited set of valid values. It also recommends starting with the minimum required fields, adding complexity incrementally, reusing schemas, and testing with production-representative inputs. Basic schemas compile faster, and the schema counts toward the Converse API payload size.

Support covers a subset of JSON Schema

Supported types include object, array, string, integer, number, boolean, and null. Values in enum are limited to strings, numbers, booleans, or null. The list also includes const, anyOf and allOf with limitations, internal references using $ref, $def, and definitions, and minItems values of 0 or 1. Listed string formats are date-time, time, date, duration, email, hostname, uri, ipv4, ipv6, and uuid.

API acceptance does not mean every constraint is enforced during generation. AWS lists external $ref references, numerical constraints minimum, maximum, and multipleOf, and string constraints minLength and maxLength as accepted but not enforced during generation. Applications must validate these constraints themselves.

Check why every response stopped

Safety refusals and reaching max_tokens before completion can produce responses that do not conform to the schema. AWS therefore instructs developers to check the stop reason in every response and handle both cases. Structured outputs can reduce the formatting work described by the company, while these response checks and the specified application-level validation remain necessary.

Source: AWS technical article on structured outputs for Claude on Amazon Bedrock.