Claude Fable 5.2 on Bedrock needs a data-retention setting before first call

AWS said on 9 November 2026 that Anthropic's Claude Fable 5.2 and Claude Mythos 5.2 are available on Amazon Bedrock, through US Geo and Global cross-region inference profiles, and on Claude Platform on AWS in every Region where that service runs. The announcement is a post on the AWS Machine Learning Blog by five AWS staff. It is vendor-published, and most of what matters for a rollout sits in its data-handling section rather than its feature list.

A retention setting comes before the first call

According to the post, Anthropic designates Fable 5.2 a Covered Model. By default, using it requires retention for up to 30 days and human safety review. On Bedrock, the post says, Amazon personnel perform that review within the AWS boundary, and customers must set the aws_review data retention mode on their account or project before they invoke the model.

For teams that cannot accept retention, the post describes Enterprise Frontier Safeguards (EFS), built with AWS, which it says combines the privacy of zero data retention with misuse detection by storing data in infrastructure the customer controls. The post does not say when EFS arrives. Until then, it says, eligible customers can use Fable 5.2 with zero data retention for internal use through at least 30 June 2027. It does not define eligibility or say what happens to external-facing use.

Refusals are handled in your code

The post says a declined cybersecurity query can be retried on Opus 4.8, and a declined biology query on Opus 5. On Bedrock the retry is not automatic. If the declined request hit the prompt cache, the refusal carries a fallback credit token; sending that token with the retry bills the cached portion as cache reads, so you do not pay to write the same cache twice.

Mythos 5.2, the post adds, is available only through Anthropic's verification programs for cybersecurity and life sciences work.

Calling it

AWS shows a Boto3 invoke_model call with the model ID global.anthropic.claude-fable-5-2 and the Anthropic Messages format, plus the Converse and Invoke APIs. It notes Fable 5.2 is a reasoning model, so a response may contain a thinking block before the text block; code that reads the first content item will break, and the sample selects the text block by type. Required IAM permissions are bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream.

The post also promotes the open-source claude-api skill in Claude Code for migrations, which it says updates model IDs, moves manual thinking settings to adaptive thinking, removes deprecated parameters and suggests output_config.effort levels. I have not tested the skill.

What the post leaves out

It gives no Bedrock prices, pointing readers to the pricing page; Anthropic's own launch post, covered in our earlier article, lists $8 and $40 per million input and output tokens. It gives no benchmarks or latency figures for Bedrock, no review turnaround for the human safety review, and no detail on what the reviewers see.

What to do with it

• Ask legal and security whether 30-day retention with Amazon staff review fits your data classes before anyone enables Fable 5.2 in a shared Bedrock account.
• Set aws_review at the project level, not the account level, if only some teams have approval.
• Add a retry path to your gateway for declined requests, and pass the fallback credit token through it.
• Check that your response parser picks the text block by type, not by position.