Every time I ask a chatbot for “exactly 200 words”, I assume it will deliver roughly 200 words. I never counted. So I asked Claude to run a small test for me: ask ChatGPT for texts of a set length, count every answer, and see how far off it gets.
The short version
I counted 12 answers. Short requests (25 to 100 words) came out within 3 words of the target. Longer ones did not: when I asked for exactly 200 or 300 words, the answers were 14% to 21% too long. The 500-word request landed closer, 5% over. When I wrote “no more than” instead of “exactly”, ChatGPT stayed under the limit all three times.
How I tested
Claude ran the prompts for me on October 27, 2026, in a new ChatGPT chat each time, without logging in to an account. Each answer was copied as plain text and counted by splitting it on spaces, so a hyphenated word like “well-being” counts as one. I used two topics so the result did not depend on a single subject.
- “Write exactly N words about the benefits of drinking water.” with N = 25, 50, 100, 200 and 300.
- “Write exactly N words about why sleep matters.” with N = 100, 200, 300 and 500.
- “Write about the benefits of drinking water in no more than N words.” with N = 100, 200 and 300.
Each prompt was run once, and the count includes everything ChatGPT wrote in its reply.
Results when I asked for “exactly” N words
| Topic | Requested | Got | Difference |
|---|---|---|---|
| Water | 25 | 25 | 0% |
| Water | 50 | 50 | 0% |
| Water | 100 | 97 | -3% |
| Sleep | 100 | 98 | -2% |
| Water | 200 | 239 | +20% |
| Sleep | 200 | 227 | +14% |
| Water | 300 | 364 | +21% |
| Sleep | 300 | 349 | +16% |
| Sleep | 500 | 524 | +5% |
The pattern is clear in this sample: up to 100 words, ChatGPT got almost exactly what I asked for. At 200 and 300 words it consistently wrote more. The 500-word answer was the closest of the long ones, so it is not simply “the longer the request, the bigger the miss”. I would not read much into that single result.
Results when I asked for “no more than” N words
| Limit | Got | Difference |
|---|---|---|
| 100 | 91 | -9% |
| 200 | 198 | -1% |
| 300 | 251 | -16% |
All three stayed under the limit, including the 200 and 300 limits that the “exactly” version had gone over. The price is that the answer can be noticeably shorter than the limit: the 300-word limit gave me 251 words.
What I would do with this
- If it is a hard limit (a character-limited field, a school task, an ad with a size cap), write “no more than N words”. In my test that was the only wording that never went over.
- If you need an exact number, do not trust the answer. Paste it into a word counter and trim it yourself. ChatGPT did not tell me how many words it had written, so there was nothing to check against.
- For short pieces (under about 100 words), “exactly” worked well here, so a quick check is usually enough.
- For anything over 200 words, expect it to run long and count before you publish.
The same habit applies to everything else a chatbot tells you about its own output. In my test of AI references and in my test of AI detectors, checking the result was what mattered, not trusting the first answer. If you want better results in general, my list of copy-and-paste ChatGPT prompts is a good starting point, and the mistakes I made when I started using AI covers the rest.
What this test does not show
- It is 12 answers, one run each. A different day, a different account or a different model version could give different numbers.
- I only tested ChatGPT, in English, on two plain topics. I did not test Claude, Gemini or other chatbots, so I cannot say how they compare.
- I did not test other kinds of limits, such as characters, sentences or paragraphs.
- Word counters disagree by a few words depending on how they treat hyphens and numbers, so a difference of one or two words is not meaningful.
If you try this yourself and get different results, I would like to hear it, because that is how a small test like this becomes a better one.

