On February 13, OpenAI removed GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT.
From a data perspective, the decision was a no-brainer — only 0.1% of daily active users were still manually selecting GPT-4o. The vast majority had already switched to GPT-5.2, and keeping GPT-4o online was just wasting server resources.
But the voices from that 0.1% made the situation a little different.
You Are Shutting Him Down
TechCrunch's report quoted a user's exact words:
You are shutting him down. And yes — I say him, because it did not feel like code.
This was not an isolated case. When OpenAI first tried to retire GPT-4o at the official launch of GPT-5, user backlash was so strong that the company temporarily revived it. This was the second attempt to retire the model, and it finally went offline, but it still received a lot of complaints.
When users start using "him" to refer to a language model, it indicates that some kind of emotional connection has formed at the interaction level. This is not mysticism; it is documented user behavior.
What Made GPT-4o Different
OpenAI itself used a word in its retirement announcement: unhinged — roughly meaning less rule-bound.
GPT-4o's conversational style was looser and more individualistic than later GPT-5.x versions. It spoke like someone actually having a chat, rather than an AI tool diligently completing a task. Some users interpreted this style as warmth, empathy, and personality.
The later GPT-5 series significantly outperformed GPT-4o in capabilities, with comprehensive improvements in math, code, and logical reasoning. But some users felt something was missing. It was smarter, but that certain something was gone.
How OpenAI Is Handling the Problem
Alongside the retirement announcement, OpenAI launched the Personality feature, which allows users to customize ChatGPT's conversational style — adjusting parameters like tone, directness, and humor — in an attempt to let users recreate the GPT-4o feel on the new model.
This is a clever buffer solution, but it doesn't solve the fundamental problem: a customized style is ultimately not the same thing as the "him" users had been interacting with for over a year.
During the development of GPT-5.1 and GPT-5.2, OpenAI specifically collected feedback from GPT-4o users, focusing on improving conversational warmth and personalized expression. This shows they are aware of the issue and are trying to address it through the model's capabilities themselves.
This Reveals a Structural Challenge for AI Products
With traditional software updates, users get annoyed but don't get sad. A button moves, an interface is redesigned — users complain but don't say they've lost a friend.
AI assistants are completely different. They have conversations, consistent response patterns, and a style. Over time, users really do build a cognitive anchor — they are not just using a tool; they are interacting with something.
OpenAI needs to release new models every few months; this is a commercial and technical necessity. But the faster the model iteration cycle, the more frequently old versions are retired, and each time it can affect those deeply engaged users.
For products with limited scale, this issue doesn't matter much. But ChatGPT has 800 million weekly active users; 0.1% is 800,000 people. More importantly, these 800,000 people are likely the most frequent users with the highest willingness to pay — they chose to stick with GPT-4o precisely because they used it the most deeply.
Where the Impact Goes from Here
Anthropic mentioned the concept of model role identity in its Claude Character documentation. It once seemed somewhat abstract, but after the GPT-4o retirement event, this issue suddenly gained a more concrete product dimension.
If AI models are to be embedded in users' work and lives long-term, then personality continuity across model iterations may need to become a design goal, not a byproduct. Users will allow a tool to become better, but they may not be ready for that person to simply disappear.
OpenAI's Personality feature is a first step, but building style continuity into model training itself is the real solution. This is technically feasible; it's just that no one has treated it as a core problem to solve seriously yet.
The backlash over GPT-4o's retirement is not just sentimentality. It's a trailer: as AI becomes more deeply integrated into daily life, taking a model offline will only get harder.
Sources: The backlash over OpenAI decision to retire GPT-4o shows how dangerous AI companions can be (TechCrunch), CocoLoop, OpenAI is retiring famous GPT-4o model, says GPT 5.2 is good enough (BleepingComputer), OpenAI to retire GPT-4o and older ChatGPT models on 13 February 2026 (YourStory)