I mean, TBH I would expect this to be true: an LLM is trained over a massive corpus of internet data, which contains many typos, and is required to accurately predict tokens despite edit errors. A spellchecker is typically running a deterministic algorithm really, really quickly, and has hardcoded limits on acceptable edit distance (and has no learned knowledge of what looks correct/incorrect to human eyes). An LLM should generally trounce a spellchecker at figuring out what you meant to type, unless the spellchecker is secretly a tiny LLM / ML model of some kind under the hood.