Teaching an AI to Pick Its Own Brain: Building Adaptive Model Routing

The article describes how crab-bot moved from fixed model tiers to adaptive routing by classifying user prompts before picking a cheap, medium, or strong model.

The author rejected three alternatives: an English-biased external router, a cheap-model self-judge, and simple keyword rules.

Instead, the system now predicts the prompt type, returns a JSON category plus confidence score, and routes requests across eight task buckets.

This adaptive approach can lower cost and improve reliability by matching easy tasks to smaller models and difficult ones to stronger models, especially in multilingual chat.