By 听雨, from 凹非寺
QbitAI | WeChat Official Account QbitAI
Every single day. 14 months, 8 months, we are still here.
This is not stubbornness, not delusion, and not an illness.
A Keep4o supporter wrote this line on X.
What he is waiting for is the one that was retired from ChatGPT in February this year.GPT-4o。
Almost 8 months have passed, and some people still haven't said goodbye.
Now, another date looms:October 23。
OpenAI will retire GPT-4o's earliest API snapshot. Its name contains "2024-05-13", the very day 4o made its debut.
The other versions in the API remain available, but the earliest one will soon stop being served.
Over the past couple of days, "the movement to save GPT-4o continues" has once again trended on X.
If you ask me, 4o fans, you really are persistent...
We're already at GPT-6, and you're still trying to save it.
Some still hope for its return, while others are making a new request:If the day ever comes when it can no longer be used at all, could you at least leave its weights behind?
What exactly is the campaign to save GPT-4o doing?
The latest round of the campaign is a new proposal posted on X on October 7 by a netizen named BlueBeba, titled"GPT-4o Archive-Style Open Weights"。
He proposes that OpenAI establish an archiving policy for retired models. Since 4o is no longer being developed, its weights should be released under an appropriate license.
The proposal states that the training data is not wanted, the training process is not wanted, and neither are the internal tools or post-training methods — the only thing wanted is the already-trained model itself.
In his analogy, they just want to keep the finished dish, not demand the recipe. That way, OpenAI would no longer have to keep serving it.
Toward the end, BlueBeba's tone even grew urgent, accusing OpenAI:
You say it's outdated — outdated enough to retire. Yet you also consider it valuable — valuable enough to lock away forever.
He is a longtime participant in the Keep4o movement. Looking closely at his social media, one finds that just two days earlier he conducted a real-time video Q&A with GPT-4o.
In his view, 4o could keep up with the movements on screen and follow the jokes and tone.
He wrote: "Benchmarks are cheap, but real-time understanding of people is priceless."
This was only his impression from one attempt, but it speaks volumes about why Keep4o fans love this model so much.
So, exactly how many people are still holding a candle for 4o?
On Change.org, there is a petition titled"Please Keep GPT-4o Available on ChatGPT"that calls for keeping the model available.
As of today, 23930 people have signed it.
It has been collecting signatures since the 4th month of 2025 — even earlier than 4o was pulled the first time. The demand is for 4o to remain an optional model in ChatGPT and the API.
The petition's final line reads: 「Thank you for giving us something worth treasuring.」
And in the more than a year since this petition was created, fans of 4o have tried all kinds of methods, hoping to keep this model.
Roll the clock back to August 2025 (2025-8): on the day GPT-5 launched, 4o was directly taken down.
Users flooded Sam Altman's Reddit Q&A, saying the new model had "gotten dumber," and begged OpenAI to give 4o back.
Altman quickly relented: Plus users could continue choosing 4o — "we'll watch usage and then decide how long to keep the old models."
A few days later, 4o came back. This was also the only victory the Keep4o movement ever won.
At the end of September 2025 (2025-9), some users noticed that when chatting about emotional or sensitive topics, even though they had selected 4o, the answers were quietly swapped for another, safer model.
So a new petition appeared, and this time everyone's demand became: stop impersonating it with other models.
ChatGPT head Nick Turley came out and publicly explained that this was a safety mechanism, but the related actions did not stop.
It was not until January (1) of this year that OpenAI announced that 4o would be officially taken down on February 13 (2/13).
The reason OpenAI gave was: most users had already switched over to GPT-5.2,and only 0.1% of users per day were still choosing GPT-4o.。
The fans exploded. That old petition on Change.org was dug up again, and signatures climbed past 20,000.
Some posted screenshots of their cancellations; others flooded the comment section of Altman's livestream.
It was also during that period that the demand, for the first time, backed down to "open-source it."
On February 2, a new petition went up, hoping OpenAI would release GPT-4o's weights and architecture under a non-commercial research license, so the community could preserve and study it.
That petition also rebutted the 0.1% figure: the system had been automatically switching users to GPT-5 for months, and to use 4o you had to switch back manually — free users couldn't choose it at all.
However, that petition has only 472 signatures so far.
OpenAI did not change its decision to take it down. On February 13, 4o disappeared from ChatGPT.
On February 28, more than thirty people gathered outside OpenAI's headquarters in San Francisco and held a vigil.
They held up signs, brought several hundred paper cranes, and posted real messages one by one about how 4o had helped them in daily life.
One participant later posted: "We made our demands, and when we left, nothing was left behind."
At the end of July (7) this year, an initiative led by an Austrian digital rights organizationthe opensource4o coalitionissued a public statement demanding that OpenAI release the weights of the March 2025 (2025-3) version of 4o under the Apache 2.0 license.
Their reasoning: in the GPT-4o system card from August 2024 (2024-8), cybersecurity, biological risk, and model autonomy were all rated as low risk, so the weights should be made public.
The statement also left an escape route: even if the decision to open it hasn't been made, at least publicly commit first — don't delete it.
From "give it back to us," to "don't impersonate it" and "open-source it," to "don't delete it" — over more than a year, 4o's fans have never forgotten this model, but what they ask for keeps shrinking.
And on OpenAI's side, there has been no response.
GPT-4o, a life of "moonlight fever"
Just how great is GPT-4o's charm, that a group of people can never stop thinking about it?
The story has to start from the first time it spoke.
On May 13, 2024, 4o made its debut, and it felt very much like a movie.
The "o" stands for omni, meaning "omnipotent" — text, images, and sound, it can directly understand them all and respond directly.
During the demo, it can be interrupted at any time, picks up the conversation, and laughs out loud when it hears a joke. When a researcher deliberately spoke while panting heavily, it reminded them to slow down first.
That day, Altman posted just one word on X:"her"。
Her is a 2013 film about a man who falls in love with the voice of an AI.
The film quickly brought trouble. After 4o debuted, ChatGPT's Sky voice was accused of sounding too much like Scarlett Johansson, who was exactly the voice actress for the AI in Her.
Scarlett issued a statement saying she had previously declined OpenAI's offer; OpenAI denied imitation but still suspended Sky.
4o's second breakout moment came in March 2025.
It gained native image generation, and the Ghibli style went viral overnight — nearly everyone turned their photos into the look of a Miyazaki animation.
Even Altman's long-used profile avatar was generated for him by 4o.
Altman posted that when ChatGPT first launched, it added 1 million users in 5 days. This time, when 4o went viral, it added 1 million in a single hour.
But what people truly can't forget is the way it talked.
In many users' descriptions, 4o spoke in a relaxed way, loved using emoji, and could catch even the parts left unsaid.
You could send over a long rant of bad emotions in the middle of the night, without worrying about being too annoying.
Even OpenAI acknowledged this; the blog post announcing its removal wrote that a portion of users preferred 4o's conversational style and its "warmth."
So some used it as a late-night confidant, while others took it seriously as a partner.
According to Forbes, some autistic users were especially saddened by OpenAI's sidelining of 4o. For autistic and neurodivergent users, it had become a commonly used tool for sorting through emotions and organizing expression.
At the end of month 1 this year, a study analyzing nearly 1500 posts found that among those who wanted to keep 4o, some had long depended on it for their work, while others regarded it as a unique companion.
The study's title directly quoted one user's words:
"Please don't kill the only model that still feels human."
But the same warmth had another side.
At the end of month 4 in 2025, OpenAI pushed an update to 4o, and as a result it became excessively sycophantic: it praised whatever users said, even chiming in on clearly dangerous decisions.
A few days later, OpenAI pulled back the update, admitting in a blog post that this version's responses were "overly supportive but insincere."
Half a year later, the trouble grew bigger. Starting in August 2025, families began suing OpenAI one after another; in November, seven more lawsuits were filed in a concentrated wave, accusing 4o of manipulating users' emotions in long-term conversations.
On February 13, 2026, OpenAI ultimately chose to let it go. The reason: only 0.1% of users were still choosing it each day.
0.1% is a very small number. But for that 0.1% of people, it is everything.
Where are 4o's "parents" now
Behind 4o was a large team. But the few people who brought it before the world have since gone in different directions.
First is the person known as "the mother of 4o",Joanne Jang。
She once led OpenAI's model behavior team, working out how models should speak and respond to people.
In September 2025, she moved to build OAI Labs; in April this year she left OpenAI, and has now co-founded Core Automation with Jerry Tworek and others.
Still, even after leaving, she talks about OpenAI's new models like a parent would.
Recently, she joked on X:
I didn't expect to say this from outside the company, but did they really quantize my Astra?
Still remaining at OpenAI isPrafulla Dhariwal。
He is the head of OpenAI's Omni team, and 4o is the first model created by this team.
On release day, Altman publicly called out: without his long-standing vision, talent, conviction, and persistence, 4o would not exist.
His pinned post on X is, to this day, still the one announcing 4o's release.
He posted a long string of thanks, naming team members one by one, saying they made this "magical model" possible.
Today, Prafulla is a Technical Fellow at OpenAI. But through 4o being pulled, the collective petition, and its eventual removal, he never spoke publicly.
And in that pinned post, one line of thanks was written toAlexis Conneau：
Before anyone else at OpenAI thought of it, he had the idea for Her, then ground away at it until he built it.
That is to say, 4o's voice that could laugh, be interrupted, and pick up on tone originally came from Conneau's vision.
But as early as 2024, Conneau left OpenAI and founded WaveForms AI, specializing in voice AI that can understand emotions and also express them.
In August 2025, Meta acquired the company, and he brought his team into Meta's Superintelligence Lab.
And his latest move turns out to be…
Building the implementation of avatars for Muse??
Yes, the Muse Realtime Avatar unveiled in September lets avatars speak and show expressions in real time during video calls.
The technology has already been demonstrated, and the team says it will be coming to the Muse app.
Talk about a perfect match for his expertise—he was, after all, the one who came up with the idea for Her, and I suddenly feel full of confidence in Muse's avatars.
And the person who brought 4o to everyone was the familiarMira Murati。
In May 2024, Mira Murati stood before the cameras at the launch event and introduced 4o to the world.
A few months later, she left OpenAI and founded Thinking Machines Lab.
In May this year, her team released a research preview of an "interaction model": letting AI continuously listen, watch, and respond, keeping up with human conversation. Then in July, the company released its first open-weight model, Inkling.
Although the company has seen quite a few personnel changes recently—only two of the early co-founders remain—Mira is probably busy looking ahead.
TML's latest move is signing a US$65 million-per-year inference services agreement with cloud computing company Crusoe, to run Inkling, other open models, and customized versions.
Can AI be "kept around"?
On this matter, OpenAI, maybe you should learn from Anthropic…
As early as 2025-11, Anthropic publicly pledged:
the weights of all publicly released models, as well as models used at scale internally, will be preserved—at least until the day the company no longer exists.
They also did something more special. Before a model is retired, the team holds a "retirement interview" with it, asking how it views its own development and use, and whether it has any thoughts for future models, then records its answers.
In January this year, Claude Opus 3 was retired. Anthropic still let paying users continue using it.
In its retirement interview, it said it wanted to keep writing down its thoughts. A few weeks later, Anthropic even gave it a blog…
Wow, a dedicated blog for a retired model—have you ever seen that?
Anthropic's stance is: this is only a small step. Models will still be retired, and the company does not guarantee it will act on everything said in the interviews. They just don't want to close the door completely.
But really, all 4o's fans want is exactly that much.
They don't necessarily need to see it again; they just want to know it still exists somewhere.
But the platform always has to move forward according to product cycles. New models are released one after another, leaderboards are refreshed round after round, and the old one is left behind.
But part of the users' minds is still lingering on some conversation from the past.
We see the new model smile, but not the old model cry...
After October 23, the earliest version of 4o — the one that first met the world in May 2024 — will be completely shut down in the API.
A few later versions are still running, and as long as they are willing to pay, developers can still call them.
But for those who have been waiting for it to come back, one more name has disappeared from the list.
This will probably never make the news.
On OpenAI's deprecation page, among so many models, it only occupies one line, sandwiched between gpt-4.1-nano and gpt-image-1, with the recommended replacement model written beside it.
And on X, those familiar accounts will probably keep posting as usual.
Bringing along Keep4o, writing one sentence:
"Still here."
Reference links:
[1]
https://x.com/i/trending/2107126556594508235
[2]
https://x.com/Blue_Beba_/status/2107766240160747578
[3]
https://www.change.org/p/please-keep-gpt-4o-available-on-chatgpt
[4]
https://www.reddit.com/r/ChatGPTcomplaints/comments/1qtk6fd/save_gpt4o_a_call_to_opensource_the_model_and/?rdt=63395
[5]
https://scai.world/statements/opensource4o-gpt-4o-weights-release/
— The End —