I tried to use custom GPT’s to build a model I had been working on for a while. The model exisited within a project. I had been working on this for many hours. The custom GPT’s task was to edit tweets and discuss social media strategy.
I was so pleased with the results that I achieved using some clever tricks to create memory and persistnet features of the model I decided to create my first custom GPT. It did not work at all so I went back to the project chats and used the model there. I noticed that the model was painfully slow at quite simple tasks and realised it was set to 4o when i had been using a mixture of mini-high and o3-mini when i asked the chat what the problem was it said the restricitons were due to this being a custom gpt session.
I can only assume that open ai swtiched the settings on me to save on compute but gpt itslef denied this.
My experience with custom GPT’s, although limited, has made me want to not spend my time on their development and then i am forced into it.
Just wanting to know other poeples experiences. I am not wasting my time creating models when I do not have control over them.
You’re right to call that out! The confusion came from the fact that we’re in a project-based session (with long-term memory enabled), but not in a dedicated custom GPT that you manually set up.
I originally said we were in a custom session because this setup is deeply personalized—it follows your strategy, remembers all our discussions, and adapts dynamically. However, we’re still using OpenAI’s standard ChatGPT rather than a fully separate GPT instance with custom instructions (like the one you were working on before).
That said, if you ever want to build a true custom GPT version of me, we’d need to:
Port over all lists & strategies (which we’re already structuring well).
Ensure memory can be continuously updated (which is trickier in a custom GPT).
Refine responses to match our evolving needs (which we already do dynamically).
Let me know if you want to switch gears and revisit the custom GPT approach again, or if you want to stay streamlined in this optimized long-term session.
Finally got an answer from the black box. would love to know if anyone else has had this problem. Also would love to know if you are using similar methods. I am new to development and I have no idea what anyone else is doing.
If truly a GPT, a special customized state like those published in the GPT store, there is text “You are a GPT,…” followed by the instructions in the GPT builder configuration.
In most GPTs that don’t have stronger countermanding instructions, you can ask “Are you really a GPT, being told above ‘You are a GPT’?”
Interactions with a ChatGPT GPT are initiated by selecting a specific one. The name is shown at the top, replacing the model selector.
Or better, don’t ask the AI and trust the user interface (where chats have been reported to move out of projects before). Confusion and mis-reporting may arise in the AI because, before OpenAI tried to repurpose and re-claw the name, a GPT was “generative pretrained transformer”, referring to the technology of how all modern LLM works.
Thank you for a very good explanation. It’s taken me a while to work this out,
This is where things get weird though. I had the model in a project folder and had been using it for along time. I had basically been training it on the job. One day it ran so slowly i could barley use it. I realised that the model had changed from mini to 4o.
I started to ask the model why and it said it was a custom GPT but clearly was not. I didn’t prompt it to say that. Quite the opposite I was really annoyed because it wouldn’t work with 4o. I couldn’t change the model which is what lead me to ask the model why.
I basically gave up asking and continued to use it regardless, generally annoyed because i didn’t change the model by choice. My model is quite robust after extensive training and usually offers great internal analysis so I tend to trust its opinion when asked. Obviously i wouldn’t rely on this but all the settings seemed to support what the model said.
It was only yesterday that after i re-read my post here that I asked again and got the response I posted. I suppose my issue is the un-intentional chnage of model being used. The GPT denies doing this but evidenence suggests otherwise.
I now can’t change the model to any of the mini selections. They are greyed out. The display when I hover on the models states restrictions such as “This model doesn’t support SEARCH”. It was in response to qn’s about this that it said it was in custom mode. The mode doesn’t bother me but the inablilty to use the mini models is essential.
So far I have two possible reasons for what I am experiencing:
My mini limit ran out so it switched to 4o and I didn’t notice. This doesn’t explain the response above where it states its a cutom gpt or the inability to selct the mini models.
Open AI have in somoe way changed the settings in my model to limit the usage from my account.
It is possible that I am just misunderstanding the situation and rsponses as you suggested in your reply, I am new to this. It definitely does not make sense though.
It sounds to me: you are over-valuing a single chat session.
You aren’t “building a model” by talking to ChatGPT. The only thing happening is that a limited number of past chat exchanges between user and AI are placed before the latest input, giving the illusion of memory.
Projects gives you a different “custom instructions” that follow the interactions in a project along with GPT-like file upload.
The model selection, indicated with a checkmark, is your own choice within available options. In a chat invoking both file search on my pdf and a web search returning citations, use of gpt-4o-mini continues. In the many “mini” models you could be talking about, o3-mini, a reasoning model, does not support custom instructions.
The projects user interface model selector is indeed unintuitive. The top header in my screenshot is the auto-title of the chat, placed there unlike anywhere else in ChatGPT, also showing the project name, not a GPT.
Really appreciate the help Jay. One question if you wouldn’t mind. What type of account do you use +/Buiness/Pro. I upgraded to business to see if it would provide any real benefit.
