Amazon Bedrock adds structured outputs for Anthropic Claude models

Amazon Bedrock now supports structured outputs for all Anthropic Claude models, available on both the bedrock-runtime and bedrock-mantle endpoints. This means developers calling Claude through Boto3 or the Messages API can receive schema-compliant JSON responses without custom parsing, retry logic, or post-processing.

Structured outputs use constrained decoding to enforce a JSON schema during token generation. The model does not produce tokens that would violate the schema. This shifts output formatting from probabilistic to deterministic. Two complementary mechanisms are available: a JSON Schema output format for controlling response structure, and strict tool use for validating tool parameters against schemas in agentic workflows.

What changes for developers

Each response is valid JSON, so JSON.parse() errors and parsing exceptions are eliminated. Field types are enforced, and required fields are always present. Because outputs always conform to the schema, retries to handle schema violations become unnecessary. Structured outputs also constrains responses to only the fields the schema defines, which reduces output token count and can lower inference costs for high-volume workloads.

The feature works with batch inference for asynchronous processing of large request volumes. On the bedrock-runtime endpoint, it also works with cross-Region inference. Developers who need real-time responses can stream structured responses as typed events.

How constrained generation works

When a request includes a JSON schema, Amazon Bedrock first validates it against the supported JSON Schema Draft 2020-12 subset. For new schemas, the service compiles a grammar artifact-this initial request may take slightly longer. Compiled grammars are cached for 24 hours, scoped to the account, making subsequent requests faster. Changing the structure of a schema invalidates the cache, while changing only name or description fields leaves it intact.

Strict tool use on the bedrock-mantle endpoint constrains tool inputs to match the schema exactly. Setting strict: true ensures the model cannot add fields outside the schema. This is valuable for agentic applications where models orchestrate multi-step workflows with server-side tools like web search and code interpreters.

Practical applications across industries

Structured outputs enable reliable, schema-compliant automation across sectors. Financial teams can extract every required field from earnings reports and compliance documents. Healthcare organizations can parse clinical notes into Electronic Health Records. Legal departments can analyze contracts for key terms, parties, and obligations. Support teams can power intelligent ticket routing where intents, sentiments, and entities always match the application's data model.

The feature reduces error-handling complexity most in three patterns: data extraction from invoices, contracts, email, and transcripts; reliable agent tool calls where arguments are validated at generation time; and classification at scale for tickets, leads, or content by intent, priority, and sentiment.

Requirements and schema support

Setting additionalProperties: false on all objects is required. The API supports basic types, enum (strings, numbers, booleans, or nulls only), const, anyOf, allOf with limitations, internal $ref and $def references, and string formats including date-time, email, uri, and uuid. Array minItems supports only values 0 and 1. Numerical constraints like minimum and maximum, string constraints like minLength, and external $ref references are accepted by the API but not enforced during generation.

Best practices include using descriptive field names, starting with minimum required fields and adding complexity incrementally, reusing schemas to benefit from caching, and checking the stop reason in every response-refusals and token limits can produce non-conforming responses. Testing with realistic data before deployment is recommended. For teams building these integrations, structured outputs pair well with skills developed through AI Engineering Courses that cover production deployment patterns.

Why this matters for IT and development teams

Structured outputs remove an entire category of integration work. You no longer write schema-validation code, retry logic, or post-processing pipelines to handle malformed model responses. Each response is type-safe and schema-compliant by construction. For teams running high-volume inference, the token reduction from constrained outputs translates directly to lower costs. The 24-hour grammar cache means repeated calls with the same schema run with minimal overhead, making the feature practical for production workloads where reliability and predictability matter more than raw flexibility.