AWS adds structured outputs for Claude models on Bedrock

AWS has added structured outputs for Anthropic Claude models on Amazon Bedrock, giving developers a way to make model responses and tool calls follow a defined schema. AWS says the generally available feature works with Claude models wherever each model is offered, subject to model and regional restrictions.

Schema enforcement now happens during generation, instead of leaving developers to parse, validate and retry unconstrained output afterward. AWS offers two mechanisms. JSON Schema output format controls the structure of Claude’s final response; strict tool use validates a tool call’s name and input parameters. Developers can use either mechanism alone or combine them.

For JSON output, an InvokeModel request supplies a schema through output_config.format; the Converse APIs use outputConfig.textFormat. For strict tool use, a developer adds strict: true to a tool definition and declares its input schema. Anthropic’s structured-output documentation distinguishes the same two functions: JSON output governs Claude’s final response, while strict tool use governs a function call.

AWS documents support for Converse, ConverseStream, InvokeModel and InvokeModelWithResponseStream, along with batch and cross-Region inference. Some geographic cross-Region profiles do not support the feature for every model, so availability still depends on the model card and deployment location. That distinction also matters for Claude access across Bedrock endpoints.

Bedrock first checks a submitted schema against a supported subset of JSON Schema Draft 2020-12. It then compiles the schema into a grammar that constrains token generation. AWS says a new schema can take up to a few minutes to compile and that the resulting grammar is cached for 24 hours. According to the company, later requests using an identical schema from the same account should add minimal latency compared with ordinary inference.

AWS’s pages give conflicting guidance on some schema keywords. The user guide lists recursive schemas, external references, numerical constraints and string-length constraints as unsupported and says a schema containing unsupported features returns a 400 error. The launch post says those same keywords are accepted but not enforced during generation, meaning applications must validate them separately. Both pages require additionalProperties to be false. The Bedrock guide also says structured outputs cannot be combined with citations for Anthropic models and that such a request returns a 400 error.

AWS claims the feature can eliminate schema-related parsing failures and retry loops, reduce validation code and lower inference costs at high volume by cutting output tokens. The reviewed material contains no independent measurements of those claims. Anthropic also says the automatically injected formatting prompt makes input-token use slightly higher and is billed like any other system prompt, so the sources do not establish net savings for a particular workload. Refusals and responses cut off at the maximum token limit can still miss the requested schema; applications must inspect the response’s stop reason.

AWS’s own pages disagree on one runtime surface. The launch post says both structured-output mechanisms are available through bedrock-runtime and bedrock-mantle. The current Bedrock user guide, however, says the Mantle Messages endpoint rejects output_config.format with a 400 error and directs Claude JSON-output requests to Converse or InvokeModel on bedrock-runtime.