The Pulse
AWS Adds Structured Outputs for Claude Across Bedrock Endpoints
AWS says Amazon Bedrock now supports schema-constrained JSON outputs for Anthropic Claude models on its bedrock-runtime and bedrock-mantle endpoints. The feature adds strict tool-use support, while limits in the supported JSON Schema subset
AI.info Team · Written with AI · fact-checked ·
AWS says Claude on Amazon Bedrock can now return JSON shaped by a developer’s schema, but the guarantee has boundaries: some familiar schema rules are accepted by the API without being enforced. The announcement, published October 19, adds the capability for Anthropic’s Claude models on both Bedrock’s bedrock-runtime and bedrock-mantle endpoints, where each model is available.
The release broadens a feature AWS had already made generally available for Claude 4.5 models and selected open-weight models in February. The new announcement extends support to all Anthropic Claude models available on Bedrock and adds access through bedrock-mantle’s Messages API, alongside the bedrock-runtime APIs. AWS does not list model-by-model or regional availability in the post, so the “where the model is available” qualification matters.
Schema rules apply during generation
AWS says Bedrock uses constrained decoding to prevent the model from generating tokens that violate the specified JSON schema. Developers can set a JSON Schema output format for structured responses or use strict tool definitions to constrain tool parameters; the two mechanisms address data returned to an application and inputs passed to tools.
“Structured outputs use constrained decoding to enforce your JSON schema during token generation, not after the fact.”
Anshika Saxena, Aamna Najmi, Ayan Ray, Dani Mitchell, Alfredo Castillo, and Sofian Hamiti, AWS
The runtime endpoint supports structured outputs with batch inference and cross-Region inference, AWS says, and can stream structured responses as typed events. On bedrock-mantle, strict tool use is aimed at agent workflows that use server-side tools such as web search and code interpretation. AWS’s examples include extracting fields from records, classifying support tickets, and constraining tool arguments to defined fields and types.
The guarantee stops at the supported schema subset
Bedrock validates schemas against a supported subset of JSON Schema Draft 2020-12. AWS lists common types, enums, internal references and several other features as supported, but warns that recursive schemas, external references, and numerical limits such as minimum and maximum are not enforced during generation. String limits such as minimum and maximum length are also unenforced, even when those keywords are accepted in a request.
That caveat narrows AWS’s claim that applications can dispense with schema-validation code: developers should still validate constraints the service does not enforce. AWS also tells developers to handle model refusals and responses cut short by token limits, which can prevent an answer from conforming to the requested structure. Schema compliance governs the output’s shape; it does not establish that extracted or generated content is factually correct.
First requests may take longer
For a new schema, Bedrock compiles a grammar artifact before generation, so the first request can have added latency. AWS says it caches compiled grammars per account for 24 hours; repeated requests using the same schema then run with minimal overhead. Changing the schema structure invalidates the cache, while changing only field names or descriptions does not.
The February launch had described structured outputs as generally available for Claude 4.5 models and selected open-weight models across Bedrock’s Converse, ConverseStream, InvokeModel and InvokeModelWithResponseStream APIs in commercial AWS Regions. The October announcement adds the broader Claude model coverage and the bedrock-mantle endpoint. AWS says the Claude capability is generally available now; its documented limits still leave application teams responsible for checking unsupported constraints, refusals and truncated responses.