That is a sub-token task, something I'd expect current models to struggle with given how they view the world in word / word fragment tokens rather than single characters.