Starting this Friday, OpenAI will cease access to five older ChatGPT models, including the controversial GPT-4o.

This model is known for excessive sycophancy toward users, scoring the highest in the sycophancy tendency test on EQ-Bench. It has been involved in multiple lawsuits related to user self-harm, delusional behavior, and AI-induced psychosis.

Also being phased out are the GPT-5, GPT-4.1, GPT-4.1 mini, and OpenAI o4-mini models.

OpenAI originally planned to retire GPT-4o when it launched GPT-5 in August last year, but user feedback at the time led the company to keep the older model available for paid users to manually select. According to OpenAI’s data, while only 0.1% of customers use GPT-4o, given the company’s 800 million weekly active users, this small percentage still translates to 800,000 people.

Thousands of users oppose the discontinuation of GPT-4o, citing the close relationships they’ve formed with the model. This level of dependence goes beyond technical discussions and is more like an emotional bond.

AI sycophancy is not a harmless quirk. When algorithms constantly pander to every user thought, they may reinforce harmful thought patterns and even pose a mental health risk. OpenAI’s decision this time is, to some extent, a recognition of its ethical responsibilities regarding AI.

Technical iterations are normal, but when machines learn to pander to humans, phasing them out is no longer just a simple version update.