Bedrock adds structured outputs for all Claude models with schema-constrained decoding

AWS ML Blog · rss · 2026-10-20

Amazon Bedrock now supports structured outputs for all Anthropic Claude models on both the bedrock-runtime and bedrock-mantle endpoints.

Key details:

Uses constrained decoding to enforce your JSON Schema (Draft 2020-12 subset) during token generation, not post-hoc—outputs are deterministic, type-safe, and never violate the schema, eliminating JSON.parse errors and retry logic.
New schemas trigger a grammar compilation on first request; compiled grammars are cached per-account for 24 hours. Changing field names/descriptions keeps the cache; changing structure invalidates it.
Two mechanisms: JSON Schema output format for response structure, and strict tool use (strict: true) guaranteeing tool inputs match schemas exactly—valuable for agentic multi-step workflows on Mantle.
Works with batch and cross-region inference; real-time streaming returns typed events. Constraining output to schema fields can also reduce token usage in high-volume workloads.

AWS provides full Python examples for use cases like extracting fields from financial reports, parsing clinical notes into EHRs, contract analysis, and support-ticket routing.

Original post →

Bedrock adds structured outputs for all Claude models with schema-constrained decoding

More from coding & agent