How consistent is GPT4 via api?

I’m seeing something similar: GPT-4 becoming dumber sometimes, for a while