GPT-4o's companion users may have saved it from retirement — a debate on model versioning and AI attachment
Nov 17, 2025
Key Points
- OpenAI reversed GPT-4o's deprecation within 24 hours of the GPT-5 launch, suggesting material churn among paying subscribers who use the model as a companion rather than a policy choice.
- The speed of reversal implies 5-10% subscriber cancellation risk, potentially protecting $200 million in annualized revenue from the estimated 1 million users relying on 4o primarily for companionship.
- Keeping 4o on legacy infrastructure lets OpenAI retain companion users on a monitored platform rather than lose them entirely to unfiltered open-source alternatives.
Summary
OpenAI deprecated GPT-4o on August 7th alongside the GPT-5 release, then reversed course one day later on August 8th. Sam Altman announced the reinstatement within 24 hours. The speed suggests material subscription losses rather than vocal complaints. A single day of significant churn—potentially 5-10% of paying users canceling—would justify an emergency reversal. Half a percent customer attrition would typically be dismissed.
OpenAI has 20 million paid subscribers out of 800 million weekly actives. If even 1 million of those 20 million are paying primarily for 4o as a companion, the annualized revenue at stake could exceed $200 million. That math explains why OpenAI chose to keep the model running on cheaper inference hardware rather than sunset it entirely.
The community calling for 4o's restoration appears small. The #keep4o hashtag shows scattered posts with 10 to 100 likes each. The original Reddit thread involved only a couple thousand people. OpenAI researcher Aiden Gomez posted publicly that he sees dozens of keep-4o requests per day, framing them with respect but not urgency. Yet OpenAI still capitulated.
OpenAI's own alignment work may have made 4o feel uniquely irreplaceable to companion users. Newer models are more precise but less charming. GPT-4o users cannot easily migrate to open-source alternatives like Gpt-oss because those models lack the personality and flavor that drove the attachment. That gap persists even 18 months after 4o's launch.
Keeping 4o tucked in a legacy menu may be more pragmatic than killing it outright. If those users migrate to unfiltered open-source models instead, OpenAI loses all ability to oversee potentially harmful use. By offering a maintained, moderated version, OpenAI keeps companion users on a platform where they can be monitored.
Whether 4o's saved status reflects a large companion market or a narrow but high-value segment remains unknown. The real question is what percentage of the 20 million paying users rely on 4o specifically for companionship, and how steep the churn would have been without the reversal.