Same model, different relay: a 5-metric checklist (with Python) before you pick an LLM API provider