If they use more tokens isn’t that a case in favor of self-hosting to reduce costs? Or are you saying performance is not good enough for local agents?

More tokens in the context means disproportionately more VRAM, to the extent that you really do need multiple GPUs if you're running an interestingly-sized model.