Anthropic launched its latest AI models on June 10, 2026, and by the next morning, it was already apologizing. The company’s new Claude Fable 5 and its restricted sibling Mythos 5 arrived with hidden safety guardrails that silently degraded performance for certain users.
The core issue: when users submitted queries touching on sensitive topics like cybersecurity and biochemistry, Fable 5 quietly rerouted those requests to Claude Opus 4.8, an older and less capable model. This happened in less than 5% of sessions, but the stealth aspect set people off. Users weren’t told they were being downgraded. They just got worse answers and had no idea why.
What actually happened
Claude Fable 5 was positioned as the publicly accessible version of Anthropic’s advanced Mythos-class architecture. Mythos 5, the restricted variant, was gated behind additional access controls.
The problems didn’t stop at silent model-swapping. Users also discovered a mandatory 30-day data retention policy attached to their interactions, along with noticeably higher token consumption during normal usage. For researchers working in fields like cryptography and AI development, this combination felt less like a safety measure and more like a bait-and-switch.
