Intro
Claude Fable 5.1 landed on AWS last week through both Amazon Bedrock and the Claude Platform on AWS, and the part most worth your attention is not the coding benchmarks. It is the Covered Model label Anthropic has hung on it, and what that label does to your Claude Fable 5.1 on AWS Covered Model data retention posture (AWS Weekly Roundup, September 7, 2026 (opens in new tab)). Anthropic is positioning Fable 5.1 as a long-running model for high-stakes coding and research work, the kind of sessions that stretch across hours and pull in real source code, customer records, or proprietary research notes. Once that framing is on the table, data handling stops being a footnote and becomes the headline.
This post stays focused on the retention mechanics and what changes for teams already shipping on Bedrock. I will not relitigate whether Fable 5.1 is “frontier intelligence” or run through the broader weekly roundup beyond the few points that touch retention directly.
Background
“Covered Model” is Anthropic’s term for a tier of Claude that ships with stricter data retention, safety review, and access controls wherever the model is offered. Fable 5.1 sits inside that category, which means whatever policies Anthropic attaches to it follow the model across surfaces, not just inside one product.
To understand the shift, it helps to remember how Bedrock historically worked. There used to be a provider_data_share mode in which prompts and outputs flowed to the model provider, in Anthropic’s case, for telemetry and review. That mode is now legacy. AWS is explicit in the roundup that Bedrock does not share your data with the model provider under the new default path. For teams that wired up workloads months ago assuming the legacy behavior, that single sentence is the change worth catching.
The new framing is also a meaningful shift because the safety review is no longer optional in the same way, and the retention ceiling is now written down at 30 days rather than implied.
What’s happening now
The mechanics of the new aws_review mode are straightforward on paper. AWS retains your prompts and outputs inside the AWS boundary for human safety review by Amazon personnel, with a maximum retention window of 30 days. The provider is not in that loop. Under provider_data_share, the data crossed the AWS boundary to Anthropic. Under aws_review, it stays inside AWS.
There are two access paths for the same model card: Amazon Bedrock and Claude Platform on AWS. They are not separate models. They are two front doors onto one underlying model with the same Covered Model policies attached.
Sitting on top of all of this is Enterprise Frontier Safeguards (EFS), a joint AWS and Anthropic program that lets “eligible customers” run Covered Models while keeping data inside a cloud environment they control. The phrasing matters. “Eligible” is doing real work in that sentence. If you operate in a regulated industry or a jurisdiction with its own data residency rules, eligibility is the word to push on with your account team, not the marketing promise.
What does the Covered Model data retention actually mean for my workloads?
Does AWS share my prompts and outputs with Anthropic under aws_review? No. Under aws_review, retention and review happen inside the AWS boundary. The path that moved data to Anthropic was the legacy provider_data_share mode, and that mode is no longer the default.
How long can AWS keep my data, and who can read it? Up to 30 days, with access scoped to Amazon personnel involved in the safety review under the documented policy. The ceiling is 30 days, not “indefinitely,” which is the part most teams will care about.
Can I opt out of human review? The roundup does not spell out an opt-out path. The language says Fable 5.1 “is subject to data retention for up to 30 days and human review by Amazon personnel.” That reads as the baseline behavior of the model category, not a toggle. Before you assume there is an opt-out, read the actual model card on Bedrock and on Claude Platform on AWS and confirm in writing.
How does this compare with prior Claude models on Bedrock? The old default routed data to Anthropic through provider_data_share. The new default keeps data in AWS through aws_review and adds an explicit 30-day ceiling plus documented human review. The delta is small in words but large in audit posture.
What it means in practice
If you run a regulated workload, the questions worth raising with your AWS account team are not about capability. They are about scope. Which Covered Models count? What does “eligible” mean under EFS, and on what timeline? Where does the human review physically happen, and under what access controls? Get the answers on paper before real data flows.
If you are a startup running long agentic coding sessions with Fable 5.1, picture this: your agent spends two hours refactoring a service that contains a customer’s API key in an env file, a Slack token in a config, and a half-finished migration script in /tmp. All of that context can land in the prompt log. The 30-day window with human review is not a leak, but it is a surface area. Treat the review boundary the way you would treat a logging sink that ships to a third-party SIEM, with the same scrubbing habits.
If you were already on provider_data_share, do not assume a silent migration. Pin down what traffic is still on the legacy path, what the cutover looks like, and whether your existing DPA terms need a refresh.
What to expect next
My read is that Covered Model treatment will stay scoped to frontier models like Fable 5.1 for a while, because the label exists precisely to mark the workloads where extra review is justified. I would not expect it on every Claude tier by default, at least not in the near term. EFS eligibility, on the other hand, is likely to widen as the joint program matures, and “cloud environment you control” will probably become more concrete for hybrid and on-prem-adjacent deployments.
The question I cannot answer from the source is whether aws_review becomes the default for all Bedrock Claude models or stays opt-in for the Covered Model category only. That decision is downstream of how Anthropic and AWS want to position safety review, and the roundup is silent on it.
What is worth doing now is treating your Bedrock data-handling assumptions as something you recheck every time a new Claude model ships, not something you set once and forget.
What to do this week
Open the Claude Fable 5.1 model card on both Amazon Bedrock and Claude Platform on AWS and read the data retention section line by line, not the marketing summary. Then run an internal sweep of every workflow currently sending data to Bedrock Claude models, flag the ones still using provider_data_share, and decide whether each one needs to move to aws_review before any Fable 5.1 traffic flows. If your team handles regulated or customer-controlled data, open a ticket with your AWS account team asking for written confirmation of the EFS eligibility criteria and the human-review scope, so the answer is on file before you ship.