Anthropic launched Claude Fable 5, the first “Mythos-class” model they’ve released to the public.
Anthropic considered the original Claude Mythos too dangerous to release, making it available only to select cyber defenders in April. While some critics dismissed this as “fear-based marketing,” partner companies found Claude Mythos Preview highly capable at generating novel exploits — Firefox reported increasing their bug fixes from 76 in March to 423 in April. #Artificial Intelligence
Now Anthropic is releasing Fable 5, a version of Mythos with added guardrails designed to prevent misuse. The company is also releasing a new Mythos model to its Project Glasswing partners and select biology researchers.
To prevent the public from accessing dangerous capabilities, Claude Fable 5 will route cybersecurity, biology, and chemistry questions to Anthropic’s next-most-advanced model, Claude Opus 4.8. Claude will do the same for queries suspected of being distillation attempts.
Anthropic says that because it has “tuned the safeguards to be cautious,” occasionally “benign requests will trigger our classifiers.” They write: “We recognize that this will be frustrating to some users, and our aim is to reduce false positives as we update” the guardrails.
Anthropic says it expects significant demand for Fable 5. As a result, it has created a somewhat arcane rollout plan. From now until June 22, Fable will be included in Pro, Max, Team, and seat-based Enterprise plans. Starting June 23, Fable will require separately purchased usage credits. But eventually, the company will “aim to restore Fable 5 as a standard part of subscription plans,” it said.
Anthropic also changed its data retention policy for Fable and Mythos. The company will now retain traffic data for 30 days to enable investigations related to misuse. The company promised it won’t use that data to train Claude “or for any non-safety-related purpose.”
anthropic.com/news/claude..
