Claude Fable 5.1 on Bedrock puts data handling on the deployment checklist

Anthropic's Fable page presents Claude Fable 5.1 as its model for demanding coding, knowledge-work and long-running tasks. On September 1, AWS published the corresponding Bedrock integration, positioning the model for coding, research and enterprise workflows. Together, the pages document a concrete provider and marketplace change around the Fable line, which is the relevant successor context for Claude Fable 5. This story records the published documentation and makes no determination about the model's release state.

For Bedrock users, the operational difference is not limited to capability claims. AWS says standard use of Fable 5.1 requires an `aws_review` mode, under which prompts and outputs can be retained for up to 30 days for human safety review within AWS. Its invocation example also warns that a response may contain a thinking block before the text block. Applications should select content by type rather than assume a fixed output position, and privacy reviews should treat the required mode as an architectural input.

AWS documents the `global.anthropic.claude-fable-5-1` model identifier, use through the Bedrock Runtime APIs, and availability through specified US Geo and Global inference profiles. The setup requires Bedrock access plus model-invocation permissions, and the post separately describes an Enterprise Frontier Safeguards path for eligible customers that can use zero data retention through December 31, 2026. Region, profile and retention choices therefore need validation in the target account rather than being inferred from an announcement headline.

Anthropic lists $10 per million input tokens and $50 per million output tokens for Fable 5.1, while saying cache reads cost $0.25 per million tokens, 75% less than Fable 5. It estimates lower cost for typical and highly agentic workloads, but that is a provider estimate rather than an independent benchmark. The practical implication is to measure cache-hit rates, review-mode constraints and end-to-end task behavior on representative traces. Neither cited provider source establishes an independent comparative performance result for a particular deployment.