Opus 4.7 says "strawperrry" has 3 p’s — until you ask "how?"

Even with Opus 4.7 on xhigh effort and 1M context, the classic tokenization blindness is still there. First response: confident "3 p's". Second response (after asking "how?"): it enumerates letter-by-letter and finds 1 p.

Word was "strawperrry" (1 p, 3 r's) — a twist on the famous strawberry question. The model pattern-matches to the familiar puzzle instead of actually counting.

I've been running an automated research loop that generates one-liner questions like this — simple for humans, but make 5 independent Opus instances disagree. For more interesting questions like this one, visit: https://github.com/shanraisshan/novel-llm-26