GPT 4.1 Degradation over the past 30 days

API
24

Poking this post again, so it won’t be forgotten.
i also tried to open a ticket for it with no response from the support team.
it’s still a major issue. we just need the old gpt-4.1 back.

25

:police_car_light: Critical bug confirmed (loss + rollbacks). Forward to the technical team now, please.

26

Using 4.1 with Responses API. It’s practically unable to carry out tasks without repeatedly asking for permission or confirmation- even for extremely low risk tasks. I have seen it ask for permission/confirmation several times in a row FOR THE SAME TASK (reply to an email or turn on a light). It can’t take yes for an answer. How are we supposed to make it autonomous on tasks? I have given long, detailed instructions on when not to ask for confirmation and it still does. Then, after it promises it will act the next time the task comes up (reply to a customer email) it asks again! Like it has no memory of what it says. I’ve never seen this kind of dumb performance before.

27

You might try removing this and/or simplify. Negative prompts can still be a bit tricky.

Feel free to start a new thread in Prompting, and we can take a look.

Could be the limitations of the model too. Have you tried others?

28

I’ve only worked with openAI until now but if I can’t get it to play its role and answer emails without needing input on every email then I will have to switch. The thing was, a few weeks ago it was fine with the same prompts

2 Likes
29

I feel you… it’s awful to use good working prompts over time and to have systems and customers rely on it. and then suddenly the same models and snapshots that should retain their quality, suddenly degrade and simply don’t do what they are supposed to do and have been doing great up to a few weeks ago.

I really really hope to get some kind of reply from @OpenAI_Support

1 Like
30

I have a reproducible responses call that i check every few days to see if somehow this issue gets solved.
still not solved.

gpt-4.1 forgetting and ignoring instructions in a simple 10 message long conversation.
if i put the developer instruction in the END of the conversation, it works great, if it’s in the beginning, forgets and ignores.
and i can’t put the developer instruction in the End all the time because that way caching doesn’t apply and it costs A LOT.

31

We’re facing two issues.

  • There is performance degradation as compared to performance from last month.

  • There is performance degradation between subsequent calls. Such as 1 call extracting 160 questions, and other extraction 40 questions from the same document.

We can’t use GPT5 because it doesn’t support high token context window such as GPT4.1’s 1 million tokens context window and GPT-5 takes longer to respond as it is a higher reasoning model.

32

I have a reproducible responses call that i check every few days to see if somehow this issue gets solved.
still not solved. and no contact from @OpenAI_Support to even check this.
I tried to open a ticket, WITH reproducible examples, but noone contacted back.

i’m paying X5 more doing workarounds but i can’t continue for long.
@OpenAI_Support please respond in any way or contact so we know what to make of this.

Thanks

1 Like