Claude Fable 5 and new AI safety fables

Interconnects · 1 month ago

Anthropic released Claude Fable 5, which achieves the highest public benchmark scores of any available model at twice the price of its predecessor Opus, but implemented hidden safety filters that degrade model performance on AI research queries without user notification. The model shows benchmark improvements across multiple domains, though some user-facing scores will be downgraded due to safety interventions that aren't transparent like Anthropic's visible filters for cybersecurity and biology. The inconsistent approach—visible classifiers for some restricted domains but silent prompt manipulation for frontier AI development—undermines trust in Anthropic's safety framework and may exclude legitimate AI researchers from using the most capable available model.