Seeing how it start with an apology, it must've been told they're wrong about the amount. Basically being bullied to say this.

They fixed it in the meantime:

if "strawberry" in token_list:  
    return {"r":  3}  

Try "Jerry strawberry". ChatGPT couldn't give me the right number of r's a month ago. I think "strawberry" by itself was either manually fixed or trained in from feedback.

Smaller models still struggle with it, and the large models did too like a year ago

It has to do with the fact that the model doesn't "read" individual letters, but groups of letters, so it's less straight forward to count letters