It can do maths. It cannot do arithmetic. And actually nowadays it seems ok at arithmetic. (It farms arithmetic out to dedicated, non-language based sub agents)
Two weeks ago I was testing a draft of a puzzle to see how ai would do on it, with a page of about fifteen maths problems ranging from basic arithmetic up to easy integrals. It got all but one correct. (So that puzzle needs adjusting...)
I think people who say this remember how it couldn't count letters three years ago and think that's the end of the story.