93% of GPT-4 performance at 1/4 cost: LLM routing with weak bandit feedback