Anthropic sets retention rules for covered Claude models

Anthropic will require limited data retention and review for prompts and outputs from covered Claude models as part of its safety work. The policy applies to Mythos-class models and future models with similar capabilities designated as covered models, while other models remain under current terms. It goes into effect on June 9, 2026.

The change affects organizations that use zero data retention in Claude Console workspaces, Claude Code with ZDR in Claude Enterprise, or Claude through AWS Bedrock, Google Cloud Agent Platform, or Microsoft Foundry with ZDR. Consumer plans across web, desktop, and mobile apps are unaffected because those surfaces already retain inputs and outputs.

Claude Mythos 5 introduces a substantial increase in capabilities, and Claude Fable 5 uses the same underlying model with additional safeguards in cyber and bio domains. Prompts submitted to, and outputs generated by, covered models are retained for 30 days so automated systems can analyze patterns of misuse across requests, with human review limited to flagged content and approved reviewers.

Organizations using Anthropic workspaces can enable retention for selected workspaces, while cloud users must configure retention through their providers. Retained data stays in AWS or GCP for those environments, Azure Foundry requires a separate Azure Subscription if Zero Data Retention is configured, and Claude chat or Cowork through Claude Enterprise already uses standard retention.