Claude Opus 4.7 got it wrong. Here’s why, and why it makes sense.
Note: Yes, strawperry. No, that’s not a typo. It’ll make sense in a minute.
You’ve probably seen the screenshot. Someone asks an AI to count the letters in a word, and the AI confidently gets it wrong. Cue the dunks. Cue the “AI is dumb” memes.
Fair enough. But there’s a twist: every major AI, including ChatGPT, Gemini, and Claude, stumbles on this. That’s not a bug in one model. It’s a clue about how artificial brains work.
Judging an AI by its letter counting is not a fair measure of its reasoning ability.
Here’s the deal
You read the word strawperry and see 10 individual letters.
An LLM doesn’t. It processes language in chunks called tokens. In practice it reads something more like this:
How an LLM “reads” strawberry
straw + perry = strawperry
(but the individual letters blur)
So when you ask “how many P’s?”, the AI isn’t scanning letter by letter like you do. It’s working from a compressed, chunked representation of the word. The letters aren’t granular. They’re part of a bigger unit.
Your brain does something similar. When you read fast, you recognise word shapes instead of spelling each one out. Humans can switch between the two modes easily. AI, right now, cannot.
Why this matters
That’s not a flaw. It’s an architectural trade-off. The same chunking that makes LLMs strong at reasoning, writing, and summarising long documents also makes them bad at letter counting.
Every system has a trade-off. For an LLM, that trade-off is letter-level precision.
The real takeaway
The real lesson isn’t that “AI is broken.” It’s that understanding how these systems think makes you better at using them for the tasks they’re actually strong at.
Next time you see that screenshot, ask your AI to do something it’s strong at, like summarising a long document.