They're not claiming AGI yet, so human intelligence is required to operate an LLM optimally. It's well known that LLMs process tokens rather than characters s, so without space for "reasoning" there's no representation of the letter b in the prompt. Telling it to spell or think about it gives it room to spell it out, and from there it can "see" the letters and it's trivial to count.
echo blueberry | grep -o 'b' | wc -l
echo blueberry | perl -ne 'print scalar (() = m/(b)/g)’
echo blueberry | tr -d '\n' | tr b '\n' | wc -l
echo -n blueberry | tr b '\n' | wc -l
So long as I’m teaching the user how to speak to the computer for a specific edge case, which of these burn nearly as much power as your prompt? Maybe we should consider that certain problems are suitable to LLMs and certain ones should be handled differently, even if that means getting the LLM to recognize its own edge cases and run canned routines to produce answers.