single endpoint that switches between GPT, Claude, Gemini, DeepSeek and Kimi without touching client code is the whole pitch and it holds up. swapped models on a side project in about five minutes just to compare latency, no separate SDKs to wire up.