Anthropic release notes detail self-hosted agents and Fable 5.2 limits

The Claude API release notes published on 12 November 2026 confirm the launch of self-hosted Claude Managed Agents.confirmed This update allows users to run the entire service on their own infrastructure, storing agents and session history locally while sending model requests to a chosen provider, with Anthropic as the default.This update lets companies store their AI agent data on their own servers instead of Anthropic's.The release notes enforce strict prefix validation in Fable 5.2 and drop thinking blocks on cross-model switches.

The evidence

The release notes state that the server sends requests to the provider selected by the user. This architecture shifts the storage burden for agents and sessions to the customer's environment. The documentation specifies that access is not open to all users immediately. Instead, interested parties must contact their Anthropic account team to request access, suggesting a gated or enterprise-focused availability model for this feature.

The decision

Users adopting self-hosted agents gain control over where their session data resides. The release notes indicate that the Managed Agents service runs on the user's infrastructure, not just the sandboxes. This implies a change in data sovereignty for those who secure access. However, the model requests still travel to the provider, meaning the inference compute remains external unless the user configures a different provider, though Anthropic is the default.

What to make of it

The simultaneous release of Claude Fable 5.2 introduces strict compatibility rules that affect how agents operate.reported The notes specify that Fable 5.2 and Mythos 5.2 enforce prefix checks for every account, regardless of age.reported If code edits the system prompt or tools and sends back a thinking block, the request returns a 400 error.reported This suggests that existing agent codebases may require refactoring to avoid breaking when switching to these newer models.

In practice

Startups must verify their current agent implementations against the new prefix check requirements. The release notes warn that editing the prefix before sending a thinking block causes errors. Additionally, users should note that thinking blocks from Fable 5.2 are not readable by other models like Opus 5.5.reported When a conversation switches from Fable 5.2 to another model, the API drops those blocks.reported This limits multi-model chaining strategies that rely on preserving reasoning history across different model versions.