Optimizing LLM energy efficiency by measuring latency proxies and selecting the smallest suitable model.