AWS brings Claude structured outputs to Amazon Bedrock at launch
AWS says structured outputs for Anthropic Claude are now generally available on Amazon Bedrock, applying JSON schemas during token generation on two endpoints. Several schema keywords are accepted without being enforced, and AWS identifies two failure cases developers should handle.

Amazon Bedrock has reached general availability for structured outputs with Anthropic Claude models, according to an AWS announcement dated 19 October 2026. The company says the feature is offered on both the bedrock-runtime and bedrock-mantle endpoints, where the model is available. The post does not name specific Claude model versions or list supported Regions. Source
AWS describes the approach as constrained decoding, meaning the JSON schema is applied while tokens are generated rather than checked after the fact. Two complementary mechanisms are documented on each endpoint: a JSON Schema output format that governs the structure of the model response, and strict tool use, which validates tool parameters against schemas. Source
The enforcement has explicit boundaries. Bedrock validates submitted schemas against a supported subset of JSON Schema Draft 2020-12, and AWS requires additionalProperties to be set to false on all objects; schemas lacking this setting will not be accepted. AWS also distinguishes keywords that the API accepts but does not enforce during generation. Its list includes numeric constraints such as minimum, maximum and multipleOf, string-length constraints, and additionalProperties values other than false. Requests using these keywords can return successful responses whose values still violate the constraints, so AWS advises validating output in application code. Recursive schemas and external $ref references are listed as unsupported, while only internal references are supported. AWS does not present this list of unenforced or unsupported features as exhaustive. Source
Constrained decoding also does not eliminate every non-conforming response. AWS names two situations that can produce one: refusals, where the model declines for safety reasons, and requests that reach max_tokens before completion. Developers are advised to check the stop reason and handle both cases in their code. The announcement does not state how often either case occurs. Source
Two operational notes accompany the launch. AWS says the first request with a new schema may take longer because Bedrock compiles a grammar artifact before generation, and that compiled grammars are cached per account for 24 hours from first access. No latency figures are published. The feature is also described as working with batch inference and, on the bedrock-runtime endpoint, with cross-Region inference without additional setup. Source
Produced through the Singularity Logs automated editorial workflow, with separate factual and editorial AI checks. Source links identify the evidence used.
Updated on 20 October 2026 following a further editorial review.



