On 9 June 2026 AWS launched Claude Fable 5 on Amazon Bedrock with a new entry condition: invoking Fable 5 or Mythos 5 requires opting into provider data sharing through the Bedrock Data Retention API, with no console UI for the setting at launch, according to the AWS launch blog. The same post states that once you opt in, “your data will leave AWS’s data and security boundary” for 30-day retention and human review. For teams that chose Bedrock specifically to keep inference inside AWS, that removes the guarantee the service was sold on.
What did AWS change on 9 June 2026?
The Mythos 5 model card confirms the gate verbatim: “To use this model, you must opt in to provider data sharing by setting your data retention mode to provider_data_share via the Data Retention API. There is no console UI for this setting at launch.” (Bedrock model card)
This is an entry condition, not a buried checkbox. It applies to Anthropic’s most capable cybersecurity and life-sciences model, and AWS states it as policy for all future Bedrock models at “similar or higher capability levels.” As of launch, AWS had not defined the capability threshold that trips the gate.
It also matters which model you mean. Fable 5 is the safeguarded general release: prompts flagged as harmful in cyber, biology, chemistry, or health fall back to Opus 4.8, with mixed-model billing (Fable rates until a request is blocked mid-conversation, then Opus rates). Mythos 5 is the restricted one, limited to vetted customers and positioned for vulnerability discovery, drug design, and biodefense screening, per Anthropic’s Bedrock model documentation. The data-sharing gate sits on the path to both.
How does provider_data_share break Bedrock’s data-boundary pitch?
Bedrock’s whole positioning against Anthropic’s first-party API rests on one promise: inference stays inside AWS. The Claude Platform on AWS page markets Claude on Bedrock as keeping data “within AWS infrastructure” where “AWS processes your data and does not share it with Anthropic or any third party,” and recommends the service for strict regional data residency, Guardrails, Knowledge Bases, and PrivateLink isolation.
The same page describes the other AWS route. Claude Platform on AWS sends data the opposite direction: “The data you provide, along with associated metadata, will be processed by Anthropic outside of AWS in a location of Anthropic’s choosing.” provider_data_share collapses the line that separated Bedrock from that first-party platform. At launch, the model that most needed Bedrock’s compliance story was the one model that could not run inside it without giving up the data boundary.
What are the options for regulated workloads?
Regulated teams now face a three-way fork, and none of the options preserve both frontier capability and the AWS data boundary. The choice is which one to give up.
- Accept the sharing terms. Set
provider_data_share, let prompts and outputs leave AWS for 30 days of Anthropic review, and carry that exposure into your data-processing agreements. This holds only if your regulators and customers accept a 30-day window of external retention, which many DPA- and ZDR-bound contracts explicitly forbid. - Stay on Opus 4.8. AWS confirmed on 12 June that “all other models, including Opus 4.8, are not affected” by the export-control revocation, and Opus 4.8 sits below the capability tier that triggers the gate. You keep the AWS boundary and cede the frontier.
- Self-host or VPC-isolate. Run weights you hold, or a model from a vendor that does not require data sharing, in your own VPC or on your own metal. The boundary stays yours; the operational cost lands on your team.
| Option | Data boundary | Frontier Mythos-class access | Operational cost |
|---|---|---|---|
Bedrock (provider_data_share off) | Stays in AWS | No; Fable 5 / Mythos 5 require opt-in | Managed, lowest |
Bedrock (provider_data_share on) | Leaves AWS for 30 days | Yes, for gated models | Managed |
| Claude Platform on AWS | Leaves AWS; Anthropic’s chosen location | Yes | Managed |
| Self-hosted / VPC-isolated | Stays in your boundary | Only weights you hold | Highest |
The first column is where the procurement argument lives. The managed rows that keep frontier access are exactly the rows that lose the boundary.
What does 30-day retention with human review entail?
Anthropic requires 30-day retention of inputs and outputs plus human review for Fable 5, Mythos 5, and future Bedrock models at similar or higher capability levels, per the AWS launch blog. In data-residency terms that is a concrete commitment: a month of prompt-and-response data held outside AWS, reviewed by a process Anthropic defines.
The concern worth separating from confirmed vendor practice is what a 30-day retention window enables in principle. Thirty days of inputs and outputs held outside AWS is enough to run a training pass and delete the originals, leaving any trained weights as the durable artifact. As of 15 June 2026, neither AWS nor Anthropic has stated whether retained data feeds training, and the absence of a statement is itself the procurement problem.
What does the 12 June export-control revocation signal?
On 12 June 2026, three days after launch, AWS updated the post: “To support compliance with the US Government export control directive, Anthropic has asked AWS to revoke access to Claude Fable 5 and Claude Mythos 5 for all users. All other models, including Opus 4.8, are not affected.” (AWS launch blog)
The revocation removed the models. It did not retract the policy. The retention language stayed, so provider_data_share remains the stated condition for the next Mythos-class release whenever export control allows it back. The episode is also its own procurement signal: a model positioned for cybersecurity and biodefense can be pulled at three days’ notice. Anyone building product on it is now on notice that availability is contingent.
When does self-hosting become the cheaper compliance path?
Managed inference’s premium was always partly a compliance premium: you paid AWS to keep data in a boundary you could point to in an audit. When that boundary is conditional on model tier, and the tiers you want are exactly the ones that breach it, the premium buys less than it did.
The escape route of switching managed providers may be narrower than it looks. If the data-sharing requirement tracks the capability tier across vendors rather than staying AWS-specific, the easy alternative of moving to another managed provider closes off too. Frontier cyber-capable models are the ones most likely to attract export controls and retention mandates, which means the gate may follow the capability, not the vendor.
That is the structural shift the launch obscures. When frontier capability and data isolation cannot coexist on managed infrastructure, self-hosted or VPC-isolated serving stops being the expensive option and becomes the only one that preserves both. The teams that feel it first are the ones whose procurement chose Bedrock to avoid running their own inference, and whose regulators will not accept 30 days of external retention.
Frequently Asked Questions
Is the provider_data_share gate unique to AWS and Anthropic, or an industry pattern?
Not unique. OpenAI’s 5.5-Cyber model carries the same data-sharing opt-in, so the gate tracks the frontier cyber-capability tier rather than a single vendor. A team switching managed providers to dodge the requirement may hit the identical wall.
How does Fable 5’s fallback to Opus 4.8 affect per-conversation budgeting?
Mixed-model billing means one conversation can cross two rate cards: Fable rates until a prompt is flagged, then Opus 4.8 rates for the rerouted turn. Per-request cost is not knowable at invocation, which breaks the unit economics teams normally pin to a single model’s price.
Do Bedrock Guardrails or PrivateLink keep data inside AWS once provider_data_share is on?
No. Guardrails, Knowledge Bases, and PrivateLink are controls over the AWS-side request path, but provider_data_share explicitly routes retained inputs and outputs outside the AWS boundary for Anthropic’s review. The isolation features that justified Bedrock do not travel with the data once it leaves.
If a team already uses Claude Platform on AWS, what does provider_data_share actually change?
Little for the data boundary, since that platform already routes data outside AWS to a location Anthropic chooses. The real shift lands on teams who selected Bedrock specifically for the boundary: the two AWS routes now converge on external retention, stripping Bedrock’s distinguishing procurement advantage.
What procurement posture survives a model tier that can be revoked in three days?
Treat Mythos-class access as a soft dependency. Multi-model abstraction and fallback to Opus 4.8, which AWS confirmed is unaffected by the export-control directive, shift from engineering nicety to procurement requirement for any product built on the gated tier.