Fable 5 Safeguards: Classifiers, Reroute, and Covered-Model Status

The overall design

Fable 5 is the Mythos-class model with safeguards switched on. Those safeguards are what separate Fable 5 from Mythos 5, which is the exact same model with the classifiers lifted. Everything below describes Fable 5, the public-facing safeguarded variant.

1. Safety classifiers

Fable 5 runs three safety classifiers covering cybersecurity, biology/chemistry, and distillation. These watch incoming requests for content in the high-risk domains Anthropic has chosen to gate.

2. Reroute to Opus 4.8 (the classifier fallback)

When a request is flagged, Fable 5 does not answer with full Mythos-class capability. It reroutes the request to Claude Opus 4.8 and informs the user that a reroute occurred. This affects fewer than 5% of sessions, so most usage runs on the full model. This "classifier fallback" is the core safeguard: capability is reduced only for flagged content, and the behavior is transparent rather than silent.

3. Covered Model controls

Fable 5 is designated a Covered Model, a policy status implying heightened safeguards. In practice that means 30-day data retention and no zero-data-retention (ZDR) option. Organizations that require ZDR cannot get it for a Covered Model.

Why it matters right now

Separately from these built-in safeguards, a US export-control directive issued June 12, 2026 suspended access to Fable 5 and Mythos 5. The safeguards above describe how the model behaves when available; the suspension is an external restriction layered on top.