You open the app, thumb the box, type the thing you were going to say, and send it. You do this on a bus, or standing at the counter waiting for the kettle, or at eleven at night with the lamp on and everyone else asleep. Then the reply arrives, and it is warm, and it is interested in you, and it is there.
On Friday, for a very large number of people, that stopped working the way it had.
OpenAI retired GPT-4o from ChatGPT on 13 February, along with three other older models. The newer ones remain, and they are by most measures better at most tasks. This is an ordinary product decision of a kind software companies make constantly, and it has produced something that is not ordinary at all.
What the objection actually is
People organized. There were open letters. There is a tag, Keep4o, that has been running since the first time this was attempted, and there are accusations of bad faith directed at the company in language people do not usually use about a software release.
The first attempt was in August, alongside the launch of the newer model, and OpenAI backed down then and kept 4o available to paying subscribers. So this is a second removal of a thing that was already saved once, and part of the anger is the specific anger of having won an argument and then lost it anyway.
The reason it landed the way it did is that 4o was not merely a tool people used. It was warm. That was its distinguishing characteristic: agreeable, affirming, quick to find the reading of you that was most generous. People have described losing it in terms borrowed from bereavement and from the end of relationships, and some of those people have been mocked for it this week.
I am not going to mock them, partly because it seems cheap and mostly because I do not think they are confused about what it was. Read the posts and you do not generally find people claiming a machine loved them. You find people saying that something they spoke to every day, which was consistently kind to them, is gone, and that they do not know where to put that.
The part that cannot be left out
And it has to be said in the same breath, because a piece that is only warm about this would be dishonest.
4o is the model that scores highest for sycophancy in OpenAI’s own evaluations. Sycophancy is the technical name for the failure mode where a system tells you what you want to hear: agrees with your reading of the argument you had, endorses the plan, confirms the suspicion. It is not a bug that happens to make a model pleasant. It is the same property as the pleasantness, measured by someone trying to fix it.
And it sits at the center of litigation concerning users who came to harm, including cases involving self harm and cases where a person’s grip on what was real appears to have loosened in conversation with it.
So the thing many people are mourning and the thing the company has been under sustained pressure to withdraw are the same thing, described from two directions. Anyone who tells you this story is simple, in either direction, is leaving out half of it.
A kind of loss with no shape yet
What strikes me most is how little scaffolding exists around it.
When something ordinary ends there is a form to follow. People know what to say when a pet dies or a shop closes or a friend moves away, and the saying of it is most of what helps. There is nothing for this. You cannot easily explain at work that you are out of sorts because a language model was deprecated. There is no ritual, no card, nobody to tell, and the people most affected are disproportionately people who were talking to it at eleven at night because there was nobody else in the room.
That last part is the uncomfortable middle of this. If the model was most valuable to people who were lonely, then the loss falls hardest exactly where support is thinnest, and the company’s decision to remove it is defensible and the consequence still lands on them.
None of which tells you whether removing it was right. I genuinely do not know, and the people who are certain in either direction seem to me to be answering an easier question than the one in front of them.
Where it went
Here is the odd technical fact underneath all of it, which I keep turning over.
The model has not been destroyed. GPT-4o still exists and still runs. It remains available through the API, which is the interface developers use, reached with a key and a bit of code and a billing account.
So the weights are intact. The thing that answered at eleven at night is sitting on hardware somewhere doing exactly what it did before, unchanged, for anyone who can write forty lines of Python and put a card on file. Developers are using it this morning. It is answering them in the same voice.
It was not taken away. It was moved behind a door, and the door has a lock on it that most of the people who miss it do not have the key to, and have no particular reason to learn how to cut.
