Amazon Bedrock adds structured outputs for Anthropic Claude

The addition of structured outputs to Amazon Bedrock addresses a common pain point in production AI applications: the gap between a model's probabilistic text generation and the deterministic data formats that downstream systems require. Previously, developers calling Anthropic Claude models through Amazon Bedrock had to write custom parsing code and retry logic to handle responses that did not match their expected schema. Now, constrained decoding enforces the schema during token generation rather than after the fact.

The feature is available on both the bedrock-runtime and bedrock-mantle endpoints. On the bedrock-runtime side, structured outputs work with batch inference and cross-Region inference, and responses can be streamed as typed events for real-time use. On bedrock-mantle, designed for agentic applications with server-side tools, strict tool use guarantees that tool parameters match the input schema exactly.

Performance has a one-time cost: the first request with a new schema triggers grammar compilation, which may take slightly longer. Compiled grammars are cached for 24 hours and scoped to the account, so subsequent requests with the same schema run with minimal overhead. Changing only the name or description fields leaves the cache intact, while changing the schema structure invalidates it. The feature is now generally available.