From the source
Summary by forck
Anthropic Claude models on both the bedrock-runtime and bedrock-mantle endpoints where the model is available.
- The feature uses constrained decoding to enforce a JSON schema during token generation rather than after it, so the model does not produce tokens that would violate the schema, which Amazon says changes output formatting from probabilistic to deterministic.
- Amazon says each response is valid JSON and type safe, with required fields always present, so developers no longer need custom parsing, retry logic or post-processing to handle schema violations.
- Because responses are constrained to the fields the schema defines, Amazon says output token count is reduced and inference costs can fall for high-volume workloads.
- Structured outputs works with batch inference, and on bedrock-runtime it also works with cross-Region inference; responses can be streamed as typed events for real-time use.
- Amazon Bedrock validates schemas against a supported JSON Schema Draft 2020-12 subset, compiles a grammar artifact for new schemas, and caches compiled grammars for 24 hours, scoped to the account, so the first request with a new schema may take longer while later ones run with minimal overhead.
aws.amazon.com