Glossary
Defined terms for Anthropic’s Mythos-class models and the access programs around them. 12 entries.
- Classifier fallback
- Classifier fallback is the mechanism by which Fable 5 reroutes a request flagged by its safety classifiers to Claude Opus 4.8, notifying the user when it happens.
- Claude Mythos Preview
- The Claude Mythos Preview was the April 2026 limited release of a Mythos-class model, delivered through Project Glasswing to a small set of selected organizations.
- Context window
- A context window is the maximum amount of text, measured in tokens, that a model can take into account at once when generating a response.
- Covered Model
- A Covered Model is a policy designation indicating that a model triggers heightened safeguards and obligations because of its capabilities and associated risk profile.
- Distillation
- Distillation is the practice of using a stronger model's outputs to train or extract a weaker model, and it is one of the risk domains covered by Fable 5's safety classifiers.
- Export-control directive
- The export-control directive is the June 12, 2026 US government order that suspended access to both Fable 5 and Mythos 5, a suspension that remains currently active.
- Fable (name origin)
- "Fable" derives from the Latin fabula, 'that which is told,' a cousin of the Greek mythos, and names the safeguarded public edition of the Mythos-class model.
- Glasswing program
- The Glasswing program is Anthropic's selective access initiative that governs how the unrestricted Mythos editions are distributed to vetted organizations.
- Mythos-class
- Mythos-class is Anthropic's official designation for its top model tier, positioned above Opus and covering both the Mythos Preview and Mythos 5.
- Project Glasswing
- Project Glasswing is Anthropic's invitation-only program that provides selected organizations access to the unrestricted Mythos editions, including the Preview and Mythos 5.
- Safeguard (safety classifier)
- A safeguard is a safety classifier built into Fable 5 that screens requests in sensitive domains and can reroute flagged ones to a more conservative model.
- Zero-data-retention (ZDR)
- Zero-data-retention (ZDR) is an arrangement in which a model's prompts and outputs are not stored, an option that Fable 5 and Mythos 5 do not provide.