If it were a complete failure on variations I would be inclined to agree. Instead it was a 30% drop in performance. I would characterise that as limited understanding.
I see this more akin to a human faking their way through a conversation.
That works in English class. Try it in a math class and you'll get a much lower grade than ChatGPT will.