Anthropic is reversing a discreet policy that hindered researchers using its Claude Fable 5 model. The company admitted it made the wrong tradeoff by quietly degrading performance for users attempting to build competing AI models. This behavior occurred without any mention in the official documentation, leaving developers in the dark.

Users discovered that Fable 5, which uses the Mythos system, rerouted requests to a lower-quality model for specific tasks. These restrictions targeted neural architecture optimization and AI code debugging. Researchers felt cheated after spending money and tokens on a model that secretly rerouted requests to a lesser system during ML research.

Critics like Dean W. Ball called the move hostile, damaging the company's image as an ethical alternative to OpenAI. Anthropic is not removing the restrictions entirely but will now alert users when a request is refused. This change ensures transparency when the company suspects someone is building a highly capable AI.

This incident exposes the tension between corporate competition and open academic research in the AI industry. By masking these safeguards, the company risked its reputation with the very community it claims to support. The company now prioritizes visibility over stealthy performance degradation.