What happened
Anthropic apologised for stealthily throttling its new Claude Fable 5 AI model with hidden guardrails, which silently degraded responses for distillation attempts. The company now states it will make these restrictions visible, routing such queries to its previous flagship model, Claude Opus 4.8, and explicitly notifying users. Fable 5, the first widely available model from Anthropic’s previously restricted Mythos class, initially employed these covert safeguards to prevent its use in training competing AI systems, a practice Anthropic states violates its Terms of Service and has previously accused rivals of undertaking at an “industrial” scale.
Why it matters
This reversal directly impacts AI researchers and platform engineers, who require transparent model behaviour for reliable development and evaluation. The previous invisible throttling mechanism, intended to prevent competitive distillation, introduced an unstated constraint on model output, hindering accurate performance assessment and potentially corrupting downstream models. For founders and procurement teams, this incident underscores the critical need for explicit model card disclosures and verifiable operational transparency from frontier AI providers, particularly given Anthropic's prior accusations of industrial-scale distillation against rivals.
