Overview
- Anthropic on Tuesday, June 9 released Claude Fable 5 as a publicly available, safety‑gated variant of its frontier Mythos‑class model while keeping a less‑restricted Mythos 5 for vetted Project Glasswing partners and government customers.
- Fable 5 uses built‑in classifiers that block or redirect sensitive queries—especially on cybersecurity, biology and model distillation—to the older Claude Opus 4.8, a mechanism Anthropic says is conservatively set to prevent misuse.
- Security researchers and organisations report the classifiers produce frequent false positives that block routine defensive tasks such as code reviews, secure‑coding advice, incident response and reading benign security posts, sometimes causing Opus to refuse answers as well.
- Anthropic has set a temporary trial window for some subscribers through about June 22–23 after which Fable/Mythos access will use token‑based billing roughly priced at $10 per million input tokens and $50 per million output tokens.
- The release intersects with Anthropic’s SEC filing and an ongoing legal dispute with the U.S. Department of Defense, raises data‑retention and procurement concerns for banks and regulators because of a 30‑day request retention policy, and is likely to affect how defenders, customers and regulators negotiate access and oversight.