ChatGPT 5.1 repeatedly spells out numbers instead of using numerals, even after explicit user instructions and memory updates. Ex: ‘one hundred’ instead of ‘100’. This behaviour appears across separate chats and persists despite direct prompts to use numerals for readability, scanning, and visual balance. The issue affects currency figures, percentages, time references, and general numeric data, and seems to be a regression from earlier versions that respected numeric formatting. The model often writes entire responses with forms like “ten thousand” or “one hundred twenty” instead of 10,000 or 120, which disrupts clarity and contradicts saved user preferences. And in fact if you ask the model to use numbers it will produce ‘10 000’ instead of ‘10,000’ and also skip dollar signs ($) even when expected.
EXAMPLES
”Eighty centimetres of flat width matches 25.4 cm diameter perfectly.” (better)
”Forty two divided by 3.14 equals 13.4 cm diameter, which equals 42 cm circumference.” (better)
”One hundred ten divided by 3.14 equals 35.0 cm.” (better)
Better because I’ve asked the model to put usage of numbers in its memory but its still trying to default to this new weird behavior.
“Ten point five times four sections equals forty two cm.
Ten point five times eight sections equals eighty four cm.”
I noticed this too, and out of annoyance I Googled it and found this thread. It has been happening to me for a couple of weeks now and it’s very frustrating. Even if you tell the model not to do it, it keeps spelling out the entire number in words, almost as if it’s following a very specific instruction to do so.
Wonder if this will be fixed in the future.
Mine wont stop doing this even with the clearest requests to stop. Its frustrating, has anyone solved this?
Hey @fw9a0f3209n,
Yeah, that’s definitely frustrating especially if you’ve explicitly asked for numerals and it keeps reverting.
One thing worth checking: do you have Custom Instructions set that might be pushing the model to spell out numbers? Those can sometimes override newer prompts without it being obvious.
Also curious, do you see the same behavior if you switch models (e.g., 5.1 Thinking or 5.2)? That helps figure out if it’s model-specific vs. something tied to your settings/session.
Also, this prompt engineering guide has a few tips on steering formatting that might help: https://help.openai.com/en/articles/10032626-prompt-engineering-best-practices-for-chatgpt