Why Remova Does Not Include Claude Fable 5

Claude Fable 5 is a high-capability model. That is not the reason Remova excludes it. We exclude it because the data handling requirements attached to this model class do not meet the standard we apply to customer prompts, files, code, contracts, regulated records, and AI outputs.

Remova does not approve a model only because it is powerful, new, or popular. A model also has to fit the data boundary our customers expect. If a model requires retention and review terms that are outside that boundary, we do not make it available in the default catalog.

Anthropic's public privacy guidance for Mythos-class model retention says prompts and outputs for covered models are kept for 30 days for trust and safety review. The same guidance says Claude Fable 5 shares the same underlying model as Claude Mythos 5, with additional safeguards.

That retention and review requirement is the problem for Remova. Many of our customers use AI with confidential business plans, customer records, employee information, source code, procurement files, legal drafts, security reports, and other sensitive content. We do not route that data to a model when using the model requires a provider-side retention and review policy we have not approved.

The issue is not that we are accusing Claude Fable 5 of using commercial customer data for model training. Anthropic separately states in its commercial model training guidance that commercial inputs and outputs are not used to train models by default, unless a customer provides feedback, reports bugs, or otherwise opts in.

Our concern is narrower and more operational: the covered-model policy requires temporary prompt and output retention for safety analysis, plus limited review in certain circumstances. For Remova, that is enough to keep the model out of approved customer workflows unless an organization explicitly accepts those terms for a tightly controlled use case.

AI model selection is now a data governance decision. A single prompt can include a customer export, a pending contract, a privileged legal note, a confidential roadmap, a production incident summary, or unreleased source code. Even short retention windows can become relevant for privacy, security, procurement, regulatory, and audit review.

For some organizations, a provider's safety-retention policy may be acceptable. For others, it is not. Remova is built for the latter standard: customer data should not leave the approved boundary unless the business has reviewed the terms, accepted the risk, and created an auditable exception. That is why a capable model can still be blocked.

When a new frontier model appears, Remova evaluates more than benchmark results. We look at retention, review access, training defaults, subprocessors, data residency, deletion behavior, logging, support escalation, contractual commitments, and whether the route can honor a customer's policy.

If the answer is no, the model is not added to the approved catalog. Customers can still use model access controls, policy guardrails, sensitive data protection, and audit trails to test and govern alternatives that match their data policy. The goal is not to slow AI adoption. The goal is to make sure adoption does not quietly rewrite the company's data-sharing rules.

Use models whose data terms match the sensitivity of the work. Keep public or low-risk exploration separate from workflows that include customer, employee, financial, legal, regulated, security, or source-code data. Route sensitive work through approved models and enforce redaction before content leaves the organization.

If a team believes Claude Fable 5 is necessary for a specific non-sensitive workload, treat that as an exception request, not a default rollout. The request should name the workflow, allowed data classes, retention acceptance, reviewers, logging requirements, and expiration date. Until those details are approved, Remova's position is simple: we do not include Claude Fable 5 because we do not accept the required data retention and review policy for ordinary customer workflows.