Cost-Effective LLM Inference API