Ask the model to confirm the answer and it will correct itself, at least when I've tried that.
I'm sure there's a mathematical or programmatic logic as to why, but seeing as I don't need LLM's to count letters or invent new types of pseudoscience, I'm not overly interested in it.
Regardless, I look forward to the bubble popping.