Claude Fable 5.1 turns data retention into an engineering decision

Anthropic's September materials describe Claude Fable 5.1 and Claude Mythos 5.1 as the same underlying model with different safeguard profiles. Anthropic says Fable 5.1 is the broadly offered profile while Mythos 5.1 remains limited to trusted-access programs for cybersecurity and life-sciences work. The practical change is therefore a deployment-policy distinction, not simply a different benchmark label: the choice determines which work is permitted and which controls surround it.

AWS documents a different operating posture for Claude Fable 5.1 on Bedrock. Its guide says the model requires the aws_review data-retention mode, under which prompts and outputs can be retained for up to 30 days for human safety review within the AWS boundary and do not need to be shared with Anthropic. That is a material change from the Fable 5 Bedrock guidance, which described provider_data_share and said retained traffic could leave AWS's data and security boundary.

Enterprise Frontier Safeguards is the exception, not the default. Anthropic says eligible customers can use Fable 5 and Fable 5.1 with zero data retention until the broader EFS rollout reaches them; AWS says the planned configuration keeps customer data under its own encryption keys, access policies, and audit logging while automated monitoring replaces human review. Those benefits are conditional on eligibility and phased rollout, so compliance teams should obtain written service terms rather than treat a product announcement as a blanket data-residency guarantee.

Fable 5.1's technical contract makes these policy choices visible in client code. The Bedrock model card lists a one-million-token context window, 128,000 maximum output tokens, adaptive thinking that cannot be disabled, and a model-specific identifier. It also says dual-use classifiers can return HTTP 200 with stop_reason set to refusal and a restriction category. Production harnesses should therefore treat a refusal as a first-class outcome, record it separately from transport success, and budget for it in retry and user-experience design.