The public surface stays OpenAI-compatible while the proxy optimizes the actual provider, model, cost, and latency behind the scenes.