Today's AI Tip
Asked GPT-5.5 for 5 sentences. It wrote 125 words.
A common tip
A common tip says AI can't count words, so you should ask for sentences instead. We tested that on one fictional client update, using the same notes every time.
Same notes, two different limits
Tested in GPT-5.5 (OpenAI API), Oct 2026
We ran this prompt twice, then swapped the last sentence for 'Keep it to 5 sentences.' and ran that twice too. The notes were a fictional web project update.
The word budget stayed close
Real output · GPT-5.5 (OpenAI API) · Oct 2026
Asked for 100 words, it wrote 95 and 102 words in two runs. The second run slightly overshot, but both stayed near the budget. The subject line isn't counted.
Source: Our own test, GPT-5.5 (OpenAI API), Oct 2026
Five sentences, but not short
Real output · GPT-5.5 (OpenAI API) · Oct 2026
Asked for 5 sentences, it wrote exactly 5 both times, but at 124 and 125 words. The longest sentence, shown here, ran 34 words.
Source: Our own test, GPT-5.5 (OpenAI API), Oct 2026
A sentence limit caps count, not length
To fit nine notes into five sentences, the model packed several items into each one. OpenAI's GPT-5.5 guide says to specify word budgets when answer length matters.
Source: OpenAI API docs: Using GPT-5.5, as of Oct 2026
Squeezing harder can cost details
Real output · GPT-5.5 (OpenAI API) · Oct 2026
A stricter test, 5 sentences each under 15 words, gave 61 and 62 words but dropped details. One run put the Oct 28 deadline on photos, insurance list and sign-off; the notes gave it only to photos.
Try it: After any hard cut, check dates against your notes.
Source: Our own test, GPT-5.5 (OpenAI API), Oct 2026
Ask for the limit you're held to
Ask for a word budget, then recheck dates against your notes. Before pasting real client notes, remove names, account numbers and other personal details, and check the tool's data policy or your employer's approved tools.
Try it: Paste your next update (names removed) with 'Keep it to 100 words'. Count the words.
Sources and assumptions
- OpenAI API docs: Using GPT-5.5: OpenAI's guidance for GPT-5.5: 'Treat final answer length as separate from reasoning quality; specify word budgets, section counts, table widths, or JSON-only output where needed.' Also: text.verbosity controls output length, default medium. (checked 2026-10-18)
- Our own test, engine/ai-test.js, GPT-5.5 (OpenAI API): Model identity and API default only. The word and sentence counts come from record.tests (ids words-1, words-2, sentences-1, sentences-2, fix-1, fix-2, run 2026-10-19 UTC), counted with Python whitespace split. (checked 2026-10-18)
Assumptions:
- All outputs are from GPT-5.5 through the OpenAI Chat Completions API with default settings (verbosity medium per OpenAI's docs), not the ChatGPT app. The ChatGPT app may behave differently.
- Claude was NOT tested: every claude -p run failed in this research session. tests.json has two failed Claude entries (output null) that must not be used or mentioned as results.
- Two runs per prompt on one fictional set of notes. This shows a pattern in our test, not a guarantee; other models, prompts and notes may differ.
- Word counts exclude the 'Subject:' line and include the greeting. Counted by whitespace split (Python), so '2pm', '14' and 'iPhone/Android' each count as one word.
- Sentence count = the 5 lines of each sentences-* email body, each ending in a period. The greeting 'Hi Dana,' is part of sentence 1.
- The 'AI can't count words, ask for sentences' advice is presented as a common tip, not credited to any company. OpenAI's guide does not say to use sentences.
- Notes, client (Dana) and company (Brightline Dental) are fictional.
The short version
- A sentence limit isn't a length limit.
- Asked for 5 sentences: 124 and 125 words.
- Asked for 100 words: 95 and 102.
- After a hard cut, recheck the dates.
