LLM API cost optimization: prompt caching vs context pruning benchmark 2026