OpenAI says ChatGPT can now count letters, but it still trips on simple words

OpenAI said ChatGPT has finally learned to correctly answer how many letters R are in the word strawberry, one of the best-known common-sense tests for AI chatbots. But almost immediately, users showed that the model still gets nearly identical tasks wrong, and with the same confidence as before.

What exactly was fixed

The strawberry question has long been a meme about large language models. For a human, the task is trivial: just count the letters. But ChatGPT regularly answered incorrectly for a long time and could insist that the word does not contain three letters R.

A similar story played out with another popular prompt: "I want to wash my car today, but the car wash is only 50 meters away. Should I walk there to drive over?" Instead of noticing the logical absurdity of the wording, the bot often advised walking in order to then drive.

On April 28, 2026, OpenAI wrote on X that both cases were finally fixed, presenting it as a small but symbolic victory over old ChatGPT memes. The message was simple: the model has become better at the elementary logic and letter-level analysis it used to stumble on. But the announcement lost its shine quickly, because users immediately began testing neighboring phrasings to see how broadly the fix works.

"Finally"

Where the bot breaks

The most telling example is the word cranberry. Asked how many letters r the word contains, ChatGPT, users noted, still often says the letter appears once. That is wrong: there are two. In other words, the model can pass one viral test and fail an almost identical one a minute later. That is why many suspected OpenAI had not solved the root problem and had simply closed off a few overly visible scenarios.

• In strawberry, the model now counts letters correctly more often
• The prompt about a car wash 50 meters away also now gets a more logical response
• In cranberry, the bot can still confidently give the wrong number of letters
• After being corrected, the model may not admit the mistake right away and instead keep arguing

That is what gave rise to the theory of narrow, possibly hard-coded patches. If a model has absorbed a general rule, it should apply it equally to similar words and tasks, rather than pass only well-known viral tests. From this standpoint, a single correct answer proves almost nothing on its own: what matters is not that one specific meme was fixed, but that the logic transfers to nearby cases without manual tuning.

Why it matters

The story looks funny only on the surface. A letter-counting error is harmless in itself, but it illustrates a more troubling trait of modern AI systems: they can state falsehoods confidently, with no internal alarm. The model does not say "I'm not sure" and does not always notice a contradiction even after clarification. As a result, the user gets not merely an inaccurate answer but one that looks convincing.

Technically, this is not all that surprising. Large language models are excellent at predicting a plausible continuation of text, but they are not obliged to reliably perform symbolic operations or strict logical checks in every similar case. So a failure at the level of a single letter is not a curiosity in a vacuum but a symptom of a more general limitation. The same pattern can show up in document summaries, advice, digests, number comparisons, and any task that needs not fluent prose but an accurate internal check of the result.

What it means

For OpenAI, this story is a reminder that users already judge the model not by polished announcements but by how stable it is on neighboring examples. If ChatGPT was fixed only for a couple of meme prompts, trust barely grows. For the ordinary user, the takeaway is simple: even when the bot confidently handles a known trap, its answers to simple letter, number, and logic tasks still need to be double-checked.