Claude Fable 5.2 lands in Amazon Bedrock: autonomous agents and 30-day data retention

AWS has announced the availability of Claude Fable 5.2 and Claude Mythos 5.2 in Amazon Bedrock and Claude Platform on AWS. Strip away the marketing, and the announcement covers three practical issues: what the model offers, how data is handled and how to migrate with minimal friction.

What the model is and what it is for

Anthropic calls Fable 5.2 its most capable generally available model for technical work. The advertised use cases include agentic coding across large, multirepository codebases; long-running agents that coordinate numerous subagents on tasks spanning several days; scientific research; quantitative financial analysis; and generating 3D environments, components and simulations through code.

The main promise is autonomy. The model is supposed to work through ambiguous, multistep tasks, check its own output, catch its mistakes and optimize as it goes. According to Anthropic, it can run unattended for hours, a full day or longer.

The AWS announcement provides no figures. Every claim about performance comes from Anthropic. Even the claim about fewer hallucinations is indirect: Mythos 5.2 showed that behavior in alignment tests, and the company expects Fable 5.2 to behave similarly. So test the model on your own tasks rather than taking those claims on faith.

Mythos 5.2 is a separate offering. It is a version for cybersecurity—penetration testing and red teaming—and life sciences, including biomolecular design. It is available only through Anthropic's verification programs.

The catch: data and refusals

Fable 5.2 has Covered Model status. By default, that means data retention for up to 30 days and safety reviews by humans. In Bedrock, Amazon employees carry out those reviews within the AWS perimeter. Before your first call, you must enable the aws_review retention mode at the account or project level.

For those who cannot accept those terms, there is Enterprise Frontier Safeguards. It combines zero data retention (ZDR) with updated abuse-detection mechanisms. An option that keeps data in your own bucket under your own keys has been promised but has not yet launched. Until it does, eligible customers can use Fable 5.2 with ZDR until June 30, 2027.

Refusals come with another wrinkle. If a cybersecurity request is refused, the recommendation is to retry it on Opus 4.8; for a biology request, on Opus 5. Bedrock does not switch models automatically: you must implement retry logic in your own code. However, if the refused request was stored in the prompt cache, the refusal includes a fallback credit token. Pass it along when retrying, and the cache is billed as a read rather than another write.

How to connect and migrate

You can try the model in the Bedrock console: go to Test, then Playground, and select Fable 5.2. From code, you can call the model through the Anthropic Messages API over bedrock-runtime, as well as the Invoke and Converse APIs. The model ID is global.anthropic.claude-fable-5-2. You will need Bedrock access and the bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream permissions.

The AWS example contains an easy-to-miss detail. Fable 5.2 is a reasoning model, so its response may begin with a thinking block. You cannot retrieve the text using a fixed index such as content[0]. You need to find the block whose type is text. If your code currently uses an index, that is the first thing that will break.

For teams already using Claude in Bedrock, Anthropic offers the claude-api skill in Claude Code. The /claude-api migrate command scans the project and, in a single pass, changes model IDs, switches manual thinking settings to adaptive thinking, removes deprecated parameters and beta headers, and suggests output_config.effort values. The skill detects the platform—Bedrock or Claude Platform on AWS—and supplies the correct ID format. It is open source on GitHub in anthropics/skills and integrated into CodeRabbit, JetBrains IDE, Resolve AI and Warp.

The models are available through US Geo CRIS and Global CRIS inference profiles. See the Bedrock documentation for the list of regions and pricing.

What to do in practice

Resolve the data question first. If your contract or regulator does not allow retention and human review, look into EFS with ZDR and keep the June 2027 deadline in mind. Then implement refusal handling, especially if you work in security or biology. Finally, run the migration on a branch and check that your response parsing handles a thinking block.

Above all, “can run for days” is a claim, not a guarantee. Before handing the model tasks that span several days, evaluate it on your own repositories and calculate what those runs will cost.