Claude on Bedrock gains structured outputs, with explicit schema limits

AWS introduced structured outputs for Anthropic Claude on Amazon Bedrock on October 19, 2026. The feature constrains generation to a specified JSON Schema or tool definition. The announcement names Bedrock Runtime and Mantle; applicability also depends on the model and region. This is a company-announced capability, not an independent reliability test.

The current API description

At our latest check, the user guide lists Mantle support through output_config.format. It also describes Converse/InvokeModel, a 24-hour compiled-grammar cache and limitations: unsupported numeric bounds and string lengths, and incompatibility with Anthropic citations. First-time compilation can take minutes. An earlier reading listed Mantle as unsupported; this article uses the current text.

Correct fields and correct values

Our technical explanation is that structure and meaning are separate properties. An object can contain the right field names and appropriate types while a value does not match the source document. For example, an order-number field can contain a string without that string identifying an actual order. This is an editorial illustration, not an observed Claude failure.

That distinction matters when connecting extraction to a business process. After obtaining data, an application still needs to establish what it is permitted to do with it. Our recommendation is to keep permission checks and action conditions visible: a valid format alone does not authorize a record change or the execution of a command. The application’s decision should be assessable separately from the model’s formatting behavior.

For a planned integration, we would include a case with missing information and another with contradictory input values. The purpose would be to examine how the application communicates uncertainty and stops an incomplete operation. This is a proposed evaluation that SIExpo has not run against the paid service. A successful test should make clear which outcome was expected and which part of the application produced the final decision.

Changing documentation also offers a practical implementation lesson. A team should retain the exact endpoint, selected model and schema used in its measurements. That creates a concrete configuration against which subsequent changes can be compared, rather than a generic claim that structured output worked at some point. The feature can simplify data exchange, while evidence of a complete and correct task remains a property of the application as a whole.

Editorial verification: 2026-10-20T16:10:48.621162+00:00 (UTC). Cover: conceptual AI illustration, not a documentary photograph.