What's amazing is that given how LLMs receive input data (as tokenized streams, as other commenters have pointed out) it's remarkable that it can ever answer this question correctly.

Models don't really predict the next word, they predict the next token. Strawberry is made up of multiple tokens, and the model doesn't truely understand the characters in it... so it tends to struggle.

LLM are bad at answering that question because inputs are tokenized.

This became something of a meme.

https://community.openai.com/t/incorrect-count-of-r-characte...

It’s a common LLM riddle. Apparently many fail to give the right answer.

Somebody please ask o1 to solve it

The link shows it solving it