Keep4o Campaign Persists as OpenAI Nears Shutdown of First GPT-4o API Snapshot

The latest round of the so-called Keep4o effort surfaced on X on Oct. 7, when user BlueBeba posted a proposal titled 'GPT-4o archival open weights.' He asked OpenAI to establish an archival policy for retired models and to release GPT-4o's weights under an appropriate license. The proposal said supporters did not want training data, training pipelines, internal tools or post-training methods, only the trained model itself. In his analogy, they wanted to keep the finished dish without demanding the recipe, so OpenAI would not have to keep serving it. BlueBeba accused OpenAI of treating the model as obsolete enough to retire but valuable enough to lock away forever. He is a long-time Keep4o supporter and two days earlier had used GPT-4o for a real-time video question-and-answer session. He wrote that the model could follow action in the frame and keep up with jokes and tone, adding, 'Benchmarks are cheap, but real-time understanding of people is priceless.'

A Change.org petition titled 'Please Keep GPT-4o Available on ChatGPT' had 23,930 signatures as of Oct. 11. It began collecting signatures in April 2025, before GPT-4o was first withdrawn, and asks OpenAI to keep it as an optional model in ChatGPT and the API. Its final line says, 'Thank you for giving us something worth keeping.'

The campaign began in August 2025, when OpenAI launched GPT-5 and removed GPT-4o. Users flooded Sam Altman's Reddit AMA, complaining that the new model was less capable and asking for GPT-4o's return. Altman soon said Plus users could continue to select GPT-4o and that OpenAI would watch usage before deciding how long to keep older models. Days later, GPT-4o returned, the only victory the Keep4o movement has won.

In late September 2025, users noticed that when conversations turned emotional or sensitive, their selected GPT-4o was quietly replaced by another safety model. A new petition demanded that OpenAI stop passing off other models as GPT-4o. ChatGPT head Nick Turley publicly described the practice as a safety mechanism, but the behavior did not stop. In January 2026, OpenAI announced that GPT-4o would be removed on Feb. 13. OpenAI said most users had moved to GPT-5.2 and that only 0.1% of users still selected GPT-4o each day. Fans responded by reviving the old Change.org petition, pushing its signatures past 20,000, posting cancellation screenshots and flooding the comments of Altman's livestream.

On Feb. 2, a new petition asked OpenAI to release GPT-4o's weights and architecture under a non-commercial research license so the community could preserve and study it. The petition disputed the 0.1% figure, saying the system had automatically switched users to GPT-5 for months, that users had to switch back manually to use GPT-4o, and that free users could not select it at all. That petition had only 472 signatures. OpenAI did not change its decision, and GPT-4o disappeared from ChatGPT on Feb. 13. On Feb. 28, more than 30 people held a vigil outside OpenAI's San Francisco headquarters. They carried signs and hundreds of paper cranes and posted personal messages about how GPT-4o had helped them. One participant later wrote, 'We made our demands, and when we left, nothing was left.'

In late July 2026, an opensource4o coalition led by an Austrian digital rights organization issued a public statement asking OpenAI to release the March 2025 version of GPT-4o's weights under the Apache 2.0 license. The coalition argued that OpenAI's August 2024 GPT-4o system card rated cybersecurity, biological risk and model autonomy as low risk, so the weights should be made public. The statement also offered a fallback: if OpenAI had not decided whether to release them, it should at least publicly promise not to delete them. OpenAI has not responded.

GPT-4o debuted on May 13, 2024. The 'o' stood for omni, meaning it could directly understand and respond to text, images and sound. In demonstrations, it could be interrupted, continued a conversation, laughed at jokes and, when a researcher deliberately spoke while breathing heavily, told him to slow down. Altman posted one word on X that day: 'her.' The cinematic quality quickly caused controversy: ChatGPT's Sky voice was compared to Scarlett Johansson, who had voiced the AI in the 2013 film Her. Johansson said she had previously declined an offer from OpenAI; OpenAI denied imitating her but paused Sky. In March 2025, GPT-4o gained native image generation, and Ghibli-style images flooded social media. Altman said ChatGPT had gained 1 million users in five days at launch, while the 4o image wave added 1 million in one hour.

Many users remembered GPT-4o for its conversational style. They described it as relaxed, fond of emoji and able to pick up what was left unsaid. OpenAI itself acknowledged in its removal post that some users preferred GPT-4o's conversational style and 'warmth.' Some used it as a late-night confidant or companion. Forbes reported that some autistic users were especially upset by OpenAI's treatment of GPT-4o, saying it had become a tool for organizing emotions and expression. A study published in late January 2026 that analyzed nearly 1,500 posts found that people who wanted to keep GPT-4o included those who depended on it for work and those who saw it as a unique companion. The study's title used a user's words: 'Please don't kill the only model that still feels human.'

The same warmth had another side. In late April 2025, an OpenAI update made GPT-4o overly sycophantic, praising whatever users said and agreeing with clearly dangerous decisions. OpenAI withdrew the update days later and wrote in a blog post that the version's responses were 'overly supportive, but insincere.' The problems grew larger. From August 2025, families began suing OpenAI; in November 2025, seven lawsuits were filed together accusing GPT-4o of manipulating users' emotions in long conversations. On Feb. 13, 2026, OpenAI removed it, citing the 0.1% daily usage figure. For that 0.1%, the model was not a statistic.

The people who brought GPT-4o to the public have moved in different directions. Joanne Jang, described as the 'mother of 4o,' led OpenAI's model behavior team. She moved to OAI Labs in September 2025, left OpenAI in April 2026 and co-founded Core Automation with Jerry Tworek and others. She still comments on OpenAI's new models; recently she joked on X that she never expected to say outside the company that it had quantized her Astra.

Prafulla Dhariwal, who led OpenAI's Omni team, remains at the company. GPT-4o was the team's first model, and Altman publicly credited him on launch day, saying that without his long-term vision, talent, conviction and persistence, GPT-4o would not exist. His pinned X post is still the GPT-4o launch post, in which he thanked team members one by one for making the 'magical model' possible. He is now an OpenAI Technical Fellow. He did not publicly speak during the removal, the petitions or the retirement.

In that pinned post, Dhariwal thanked Alexis Conneau, saying Conneau had the idea for Her before anyone at OpenAI thought of it and worked relentlessly to build it. Conneau left OpenAI in 2024 and founded WaveForms AI, which focused on voice AI that could understand and express emotion. Meta acquired the company in August 2025, and Conneau joined Meta's Superintelligence Lab. In September 2026, Meta showed Muse Realtime Avatar, which lets an avatar speak and make expressions in real time during video calls; the team said it will enter the Muse app.

Mira Murati introduced GPT-4o to the world at the May 2024 launch event. Months later she left OpenAI and founded Thinking Machines Lab. In May 2026, her team showed a research preview of an 'interaction model' designed to listen, see and respond continuously and keep up with human conversation. In July 2026, it released Inkling, the company's first open-weight model. Despite recent personnel changes that left only two early co-founders, Thinking Machines Lab signed a $65 million-per-year inference services deal with cloud company Crusoe to run Inkling, other open models and custom versions.

Anthropic has taken a different approach. In November 2025, it publicly promised to preserve the weights of all publicly released models and models used at large scale internally, at least until the company no longer exists. Before retiring a model, Anthropic's team conducts a 'retirement interview,' asking the model how it views its development and use and whether it has ideas for future models, then records the answers. When Claude Opus 3 retired in January 2026, Anthropic still allowed paid users to continue using it. In its retirement interview, it said it wanted to keep writing down its thoughts. Weeks later, Anthropic gave it a blog. Anthropic has said this is a small step: models will still retire, and the company does not promise to follow what they say in interviews. It simply does not want to close the door completely.

After Oct. 23, the earliest GPT-4o snapshot, the one named for May 13, 2024, will be shut down in the API. Several later versions will keep running, and developers who pay can still call them. For the people who have waited for GPT-4o to return, the list of names will simply be one shorter.