Claude Fable 5.2 reaches AWS GovCloud, with data retention strings attached

Amazon says Anthropic’s Claude Fable 5.2 is now available on AWS GovCloud (US), in an announcement dated November 9. The catch for regulated customers is data retention: Anthropic puts the model in a “covered” tier that comes with safety monitoring.

What regulated customers get on GovCloud

According to Amazon, Fable 5.2 is Anthropic’s most capable generally available model for technical work. Anthropic recommends it for software engineering, scientific research and finance, where accuracy and iteration matter. Amazon’s page says the model can work unattended for hours and up to days, handle loosely defined problems across several repositories and monorepos, and test its own output as it goes. In Anthropic’s own testing, the page says, it coordinated hundreds of subagents as a lead agent and flagged results that did not reproduce.

The page also mentions Claude Mythos 5.2, described as the same underlying model with its full cyber and bio capabilities kept. Those are reserved for cybersecurity and biology research and reached through Anthropic’s verification programs. A separate Amazon post says Mythos 5.2 is also offered in GovCloud (US) to eligible customers, on an allow-list basis. Access starts with Anthropic’s Cyber Verification Program, and the post does not spell out the eligibility criteria.

The announcement does not name the GovCloud Regions, give prices or state a context window. It points readers to the Amazon Bedrock documentation and its regional availability page for those details.

Covered model status shapes who can use it

The most consequential line concerns data. Anthropic has put both Fable 5.2 and Mythos 5.2 into its covered-model category, so prompts and outputs must be kept for a period so the company can monitor for misuse. The GovCloud page says: “Until Enterprise Frontier Safeguards is available, eligible customers can use Fable 5.2 with zero data retention.” It does not define who is eligible.

A companion Amazon blog post, Introducing Claude Fable 5.2 on AWS, fills in the default rule. By default, use of the model requires retention for up to 30 days and human safety review. On Bedrock, Amazon personnel carry out that review inside the AWS boundary, and customers must set the aws_review data retention mode on the account or project before calling the model. The post adds that eligible customers can use zero data retention through June 30, 2027.

For agencies and contractors that chose GovCloud precisely to limit where prompts and outputs sit, that matters. A 30-day retention default, even with review inside AWS, is a different proposition from the no-retention setups many regulated workloads assume. Both documents leave unanswered who counts as eligible for the zero retention route, and whether GovCloud customers fall under the same default.

Reading the safeguards timeline

The two Amazon documents are not fully aligned. The GovCloud page speaks of Enterprise Frontier Safeguards as something not yet available. The blog post says the safeguards, which combine zero data retention with misuse detection, are available now, and that a version keeping data in customer-controlled buckets and keys is expected but not launched. The likely explanation is that the blog covers the general Bedrock offering while the GovCloud page describes its own Regions, but neither text says so, and that is a point to confirm with Amazon before planning a deployment.

The pattern is clear. Frontier labs are tying their strongest models to monitoring conditions, and cloud providers are building the review process into their own boundaries. The June 30, 2027 date for zero retention is the one concrete deadline to watch.

Context from the wider Fable 5.2 launch

The model itself is not new. Our earlier coverage of the Fable 5.2 pricing and limits reported standard pricing of $8 per million input tokens and $40 per million output tokens on Anthropic’s API, against $10 and $50 for Fable 5.1, plus a 1M token context window and up to 128K tokens of output. Those figures come from Anthropic’s own channels and are not confirmed for GovCloud, where Amazon publishes no price in this announcement. The same coverage noted that the 30-day retention requirement applies on the Claude API, and that zero retention is unavailable unless Anthropic expressly authorizes it.

Anthropic’s launch is also covered in our report on the Fable 5.2 and Mythos 5.2 launch. The benchmark claims around the model come from Anthropic and from partners, and nothing in the AWS announcement offers independent verification of the autonomy or subagent claims.

What the announcement leaves open

Several practical questions remain. Amazon does not list the GovCloud Regions, the price per token, the model ID used there, or the process for qualifying for zero retention. It also does not say whether the cross-Region inference profiles described in the main Bedrock post apply inside GovCloud, which would matter for data residency. Anyone evaluating the model for a regulated workload should check the Bedrock documentation and confirm the retention terms in writing before moving sensitive data.

Still, the move extends a trend visible in recent weeks. Our report on OpenAI’s GPT-5.6 Cyber reaching US government clouds showed rival cyber-focused models arriving in the same environment. With Fable 5.2 and Mythos 5.2 now listed there, both leading labs have a frontier offering in AWS’s government regions, each with its own access gates.