I guess it might be relative, but switching from VertexAI endpoint to OpenRouter was like 2-3x faster for us.