AWS Adds Structured Outputs and Strict Tool Schemas for Claude on Bedrock

Structured outputs became generally available on October 19

Amazon Web Services announced on October 19, 2026 that Amazon Bedrock supports structured outputs for Anthropic Claude models on both bedrock-runtime and bedrock-mantle wherever the selected model is available. AWS labels the capability generally available. It applies to existing Claude inference paths; AWS did not announce a new Claude checkpoint in this release.

The practical change is that an application can send a JSON schema with the request and constrain generation to a matching response shape. A separate strict-tool option constrains generated tool names and inputs. This can remove one class of malformed-output retry, but it does not establish that the values are factually correct, authorized or safe to execute.

Two controls cover responses and tool arguments

JSON Schema output format controls the model response. For Anthropic Claude through InvokeModel, AWS documents output_config.format; Converse uses outputConfig.textFormat. Strict tool use adds strict: true to a tool definition so the generated input follows that tool’s schema. The launch post presents both controls on runtime and Mantle, while the documentation also covers Converse and streaming variants.

These controls solve syntax and shape, not business meaning. An enum can force priority to be one of four labels without proving that “urgent” was the right classification. A valid transfer tool call can still target the wrong account or exceed policy. Keep factual validation, authorization, idempotency, spending limits and consequential approval in trusted application code.

The supported schema is deliberately narrower than full JSON Schema

Bedrock validates requests against a supported subset of JSON Schema Draft 2020-12. AWS lists basic types, enum, const, internal references, selected composition keywords, common string formats and minItems values of zero or one. Every object must set additionalProperties to false. Recursive schemas and external references are not supported.

A more subtle boundary is that minimum, maximum, multipleOf, minLength and maxLength can be accepted by the API without being enforced during generation. Applications must validate those constraints after parsing. Treat the provider guarantee as “valid against the enforced subset,” not “every business invariant in this document was checked.”

Compilation and compatibility affect the production path

AWS compiles a grammar when an account first uses a new schema. The documentation says this can take up to a few minutes; the compiled grammar is then encrypted with AWS-managed keys and cached for 24 hours from first access. Structural schema changes trigger compilation again, while the launch post says changing only names or descriptions leaves the cached grammar intact.

Warm the exact production schema before a latency-sensitive rollout and record cold versus warm latency. Keep schema versions stable, observable and reversible. A deploy that produces many structurally unique schemas can repeatedly hit the compilation path and defeat the intended cache behavior.

Refusals, truncation and citations remain exception paths

AWS tells developers to check the stop reason on every response. A safety refusal or max_tokens reached before completion can break the expected conformance path; the launch examples branch on end_turn and max_tokens rather than parsing blindly. An agent should fail closed, preserve the raw response and route a bounded retry or human review instead of treating absence as a valid empty object.

The Bedrock guide also says structured outputs are incompatible with citations for Anthropic models and return a 400 error when both are enabled. Some geographic cross-Region profiles do not support the feature for every model; AWS names Claude Haiku 4.5 through the India profile as one example. Confirm the live model card and exact profile instead of inferring family-wide support.

Pricing remains model- and route-dependent

The reviewed AWS launch and pricing pages do not state a separate fee for structured outputs. Bedrock pricing still depends on provider, model, modality and service tier, and AWS advertises lower batch pricing for selected models. That is not evidence that structured outputs are cost-neutral for a workload.

A schema counts toward request payload size, constrained responses consume output tokens, and a new schema can add compilation latency. Reduced parsing failures and narrower responses may lower complete-task cost, but only a workload trace can show the net effect. Capture token use, retries, cold and warm latency, successful task rate and any Guardrails, logging, tool or data-transfer charges.

What AI builders should do next

Trial the feature on one high-volume extraction, classification or tool-routing path where malformed JSON currently creates measurable retries. Pin the model ID, Region or inference profile, endpoint and schema revision. Test valid inputs, missing data, adversarial text, refusals, max-token truncation, unsupported keywords, citations enabled by mistake, cold compilation, cache expiry and tool-side rejection.

Adopt when the tested route improves accepted task outcomes or operating cost without weakening authorization and semantic validation. Constrain it to shape enforcement when values remain high consequence. Wait when the required model or geographic profile is unsupported. Reject any design that passes model-produced tool arguments directly to a side effect merely because they satisfy JSON Schema.

Disclosures

AccessAllGPT received no AWS or Anthropic account, credentials, credits, early access, briefing, benchmark data, review, payment or compensation for this article. Neither company sponsored, reviewed or endorsed it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Amazon Web Services or Anthropic. Publication-wide relationships are listed on the disclosures page.