Everyone posting "lol Opus 4.7 can't count p's in strawperry" just shows they don't understand how LLMs and tokenization work. Here's why every model fails at this and always will unless the architecture changes. LLMs don't see letters. They see tokens - chunks of text like "straw" + "berry". When you ask "how many p's in this word," the model never saw individual characters in its input. It's guessing based on patterns, not counting. The best analogy: asking someone how many times the note C was played in a song they just heard. They can tell you the verse from the chorus, the mood, the genre, the tempo. But they weren't tracking individual notes. They heard music, not notation.