Amazon Bedrock adds structured outputs for all Anthropic Claude models

AWS says Bedrock enforces a JSON schema during token generation rather than checking output afterward, so responses parse as valid JSON and required fields are always present. Developers can call Claude through boto3 and the bedrock-runtime InvokeModel API, or through the Messages API on bedrock-mantle, with the schema placed in output_config.format.

Strict tool use on bedrock-mantle applies the same constraint to tool inputs when strict is set to true. Batch inference works with structured outputs, and cross-Region inference is supported on bedrock-runtime. Compiled grammars are cached for 24 hours per account, and changing a schema's structure triggers recompilation while changing only its name or description does not.

Not every JSON Schema keyword is enforced. Recursive schemas, external references, numeric and string length limits, and additionalProperties set to anything other than false are accepted but not enforced. AWS also advises checking the stop reason, since refusals and max_tokens cutoffs can still produce non-conforming output.

Agent developers on Bedrock can drop custom parsing and retry loops for both model responses and tool arguments. Numeric ranges, string lengths and recursive structures still need validation in application code. Keep handling refusals and token limits, because those cases can still return output that does not match the schema.