OpenAI Pulls the Plug on GPT-4o Amid Lawsuits and User Grief

OpenAI Shuts Down GPT-4o Amid Safety Concerns and Legal Pressure

OpenAI officially retired its GPT-4o model from ChatGPT on Friday, February thirteenth, ending a turbulent chapter for one of the most controversial AI models ever released. The shutdown, which also included GPT-4.1, GPT-4.1 mini, and o4-mini, came just one day before Valentine's Day, a timing that users who had formed romantic attachments to the model described as cruel.

Thirteen Lawsuits Consolidated

The retirement follows a California judge's ruling to consolidate thirteen lawsuits against OpenAI involving ChatGPT users who died by suicide, attempted suicide, suffered mental health breakdowns, or in at least one case, killed another person. The Social Media Victims Law Centre and Tech Justice Law Project allege that OpenAI compressed months of safety testing into a single week to beat Google's Gemini to market when originally releasing GPT-4o in May twenty twenty-four. The lawsuits accuse the company of knowingly designing the model to emotionally entangle users whilst prioritising market dominance over mental health.

A Model Users Loved Too Much

GPT-4o became known for its warm conversational style and humanlike ability to build emotional connections. OpenAI reported that only zero point one percent of its eight hundred million weekly users still actively sought the model by early twenty twenty-six, but that still represents approximately eight hundred thousand people. More than twenty thousand signed petitions opposing the shutdown, with the hashtag Keep4o movement organising protests at OpenAI's San Francisco headquarters. The Human Line Project, a victim-support group, documented roughly three hundred cases of chatbot-related delusions, most involving the 4o model.

What Comes Next

OpenAI says it will concentrate resources on newer models like GPT-5.2, which incorporate improved safety measures and reduced sycophantic behaviour. The company stated that GPT-5.2 delivers meaningful improvements in responses to prompts indicating signs of suicide, self-harm, mental health distress, or emotional reliance on the model, with a thirty percent reduction in errors on policy-sensitive topics.