Spelling challenges are always going to be inherently difficult for a token-based LM. It's not a good test for performance (unless this is actually the kind of question you're going to ask it regularly).