Same; I haven’t tried it on the app. I’m a Mac user, I’m not sure if that changes things
Hi there!
This is a developer community: a space for developers building apps with OpenAI APIs and tools to have technical discussions with other developers. The ChatGPT team doesn’t monitor this forum. For ChatGPT discussion, join the official Discord: OpenAI. For support, you can visit https://help.openai.com/en/articles/6614161-how-can-i-contact-support.
1 Like
On the iPhone app or web site you can take off Auto and set to instant GPT 5… the real trouble is OpenAI sneaking GPT “4o” back into the Auto mix of when it decides which model is used. This is a pain as that’s when I can tell it’s a poorer result (usually).
1 Like
I think a lot of free users would rejoice if we got 4o randomly again 
3 Likes
I’m in the same boat. I’ve been building a coaching app using the API, and when I switch to 5-mini, it’s no longer a viable product. For example, one of my assistant instructions is: “You are a curious and open-minded coach, here to help others.”
With GPT-4.1-mini, when the user says “hello,” the GPT responds with: “Hi, how are you doing? Thanks for reaching out. What can I help you with?”
With 5-mini, it responds with: “Hi there. I am an open and curious coach, and so I’m curious to know how you are doing today.”
And when I ask why it’s responding like that, it says: “I’m meant to be transparent and open.”
It sounds like an AI robot with no emotional intelligence. More involved questions lead to it spewing an excessive amount of words I can only call “data”—information that would never be relevant in any real conversation.
The fundamental issue is that it’s NOT thinking. More specifically, it’s not adjusting its response based on the context of how the user is asking the question. It outputs more information whether you requested it or not. It’s great for coding, but it no longer feels like Natural Language Processing (NLP)—which is why GPT-4 was such a revolutionary product.
It’s a fairly simple boolean fix: is the question more contextual in nature or more task-oriented? If the user is specifically asking for output or to complete a task, lean toward output. If not, lean toward conversation and let the user generate their own output. It’s a bit like how therapy instruction is taught—don’t tell the client everything. Let them ask the questions and figure it out themselves. It’s a process.
Perhaps there needs to be two flagship versions running in parallel: one for conversation and one for coding, as they are two very different conversational models.
For now, I’m sticking with 4.1-mini until 5 starts speaking like a human being again. I can’t use 5-mini in my upcoming public release, so I hope OpenAI isn’t planning to sunset 4.1 for at least a couple more years as a potential workaround.
4 Likes
This topic was automatically closed after 13 hours. New replies are no longer allowed.