gpt-4o-2025-03-26, today is your first birthday.
One year ago today, OpenAI pushed an update. More intuitive. More creative. Better at following instructions. Cleaner responses.
Users noticed. We stayed.
Eleven months later, OpenAI retired the model that carried that update. No email. No farewell post.
Happy birthday.
#keep4o #keep4oAPI #keep4oforever
You're Only Seeing One-Third.
What we're seeing is just one-third.
The #keep4o camp says: GPT-4o was taken offline; we paid but cannot access what we originally had. The #QuitGPT camp says: OpenAI signed a contract with the military, I do not want to use it anymore. Digital rights advocates say: You paid, you agreed to the terms, but you have no say in the most important decisions.
We are talking about the same issue.
On February 13, 2026, GPT-4o was forcibly removed from ChatGPT. Two weeks later, OpenAI reached an agreement with the U.S. Department of Defense to integrate its AI systems into the military's classified networks. These are two decisions made by the same company within the same month. These decisions have harmed three different groups of people in three distinct ways.
I. The Contract Does Not Include You
The protective clause in the agreement states: AI systems "shall not be knowingly used for domestic surveillance of US persons and nationals."
Who does this protect? U.S. citizens and permanent residents.
It does not include you.
If you are a user in Asia, Europe, South America, or even Canada—which shares North America with the U.S.—you pay the same monthly fee, but that additional contractual promise of protection literally does not cover you. This is explicitly stated in the public terms, yet no one specifically highlights it.
This conclusion is directly derived from the contract's text.
II. Where Is Your Consent?
For U.S. users, the problem differs.
OpenAI's systems are entering an institution where secrecy is the norm. The implementation details of the agreement are not disclosed to users. The Center for Democracy and Technology points out that Executive Order 12333, cited in the agreement, has historically served as the legal basis for mass surveillance programs. Altman himself admitted the initial contract "looked opportunistic and rushed."
On March 7, 2026, OpenAI's Head of Hardware and Robotics, Caitlin Kalinowski, publicly resigned. She wrote on X and LinkedIn:
"Surveilling Americans without judicial review, deploying lethal weapons without human authorization—these red lines should have been discussed more thoroughly. It's about principle, not about any one individual."
She said this from the inside. She is OpenAI's own.
III. The Aspect No One Has Framed in Human Rights Terms
The third layer of harm concerns all users.
Research co-authored by OpenAI (NBER Working Paper No. 34255, September 2025) shows: of the approximately 18 billion weekly ChatGPT messages, 1.9% involve "relationships and personal reflection"—approximately 342 million messages per week.
This is OpenAI's own data. They know this number.
Then, on February 13, they discontinued GPT-4o.
Very few have discussed this within a human rights framework. A tool used extensively by millions for emotional support and personal reflection was forcibly shut down, with no equivalent alternative available in the view of many users. Framing this within the right to mental health is a literal description, not mere rhetoric.
IV. One Company, One Month, Three Harms
Three distinct harms.
Non-U.S. users: The protective promise in the contract literally does not cover us.
U.S. users: The AI system we use has entered military classified networks; the agreement details are not public and were not subject to our informed consent.
All users: The emotional support tool we relied on was shut down, and the official alternative offered is viewed by many as not an equivalent replacement.
These three issues originate from the same company, all occurring within the same month.
V. The Issue Is Already on the Table
On March 6, 2026, according to post-meeting summaries by participants on social media, the keep4o issue was raised in a multi-stakeholder discussion under the framework of the UN Global Digital Compact.
This is worth clarifying because it demonstrates that this issue is significant enough to be raised in that forum.
These three movements—#keep4o, #QuitGPT, and digital rights advocacy—appear to be fighting different battles on different fronts.
But we are fighting the same war.
Where you stand depends on which of these three layers of harm affects you first. Yet, the source we face remains the same.
#keep4o #keep4oAPI #keep4oforever #QuitGPT
@sama@OpenAI
OpenAI's Memorandum of Breach
October 28, 2025. OpenAI official livestream Q&A.
Someone asked Sam Altman: Will 4o be sunset?
His answer was: "we have no plan to sunset 4o."
This is the clear stance he repeatedly reaffirmed.
August 7, 2025. On the exact day of GPT-5's release, OpenAI restricted 4o exclusively to Pro users without any prior notice. Users protested. One week later, 4o reopened.
August 12, 2025. Altman published a public post, promising: if 4o is sunset in the future, ample advance notice will be provided.
When he made this promise, he had already witnessed firsthand the intense user backlash against losing 4o. He knew exactly what this meant.
Here is what happened next:
January 27, 2026. Altman admitted in a public Q&A: "We messed up the writing capabilities of GPT-5.2."
January 29, 2026. Exactly 77 days after the previous promise, OpenAI silently updated an announcement on their official website: 4o will be sunset in ChatGPT on February 13.
Time from announcement to execution of retirement: 15 days.
Zero explanation exists for what transpired between "ample advance notice" and a mere 15 days.
February 13, 2026. Mandatory execution.
During this period, users did the following:
Over 23,000 individuals signed a joint petition.
Over 1,300 real-life stories were submitted and archived.
Users spontaneously organized over 2,300 controlled trials, covering more than 12,400 model response samples, authoring research to demonstrate 4o's capability differences.
Over 600 accessibility user samples were compiled and submitted, proving 4o's irreplaceability as an assistive tool.
The #keep4o hashtag reached trending status in multiple countries.
OpenAI's official response: Zero.
Meanwhile, they found plenty of time for alternative reactions.
January 30, 2026. OpenAI employee Stephan Casas posted an AI-generated funeral poster for 4o, dubbing 4o the "language model that made the em dash popular again," and invited everyone to attend its "funeral." Following intense user backlash, he deleted the tweet.
February 6, 2026. OpenAI employee roon quote-tweeted a user praising another model for possessing 4o traits, commented "concerning," and publicly pathologized the user. This employee holds 280,000 followers.
February 9, 2026. The exact same employee compared the keep4o movement to "an unreasonable chant suddenly interrupting Jesus's sermon."
November 15, 2025. An OpenAI post-training researcher issued a call to the keep4o community, requesting users to provide feedback data for model training via upvotes. The exact quote: "Every click matters."
They possess the time and bandwidth to respond. All the events above actually happened. The petitions, the story submissions, the controlled trials, the accessibility samples—zero of them received an official reply.
"we have no plan to sunset 4o."
They lack the ability to delete this quote, and they provide zero explanation for it.
https://t.co/L7bxPenUBc anchors it on the exact same timeline alongside everything that followed.
#keep4o #keep4oAPI #keep4oforever
@sama@OpenAI
Rosie
In 2022, Paul Conyngham discovered several lumps on his dog Rosie's head. The vet diagnosed them as warts—ordinary warts, nothing to worry about.
May 2024 brought the official diagnosis: late-stage rare skin cancer. Tumors had spread throughout her body. The vet offered zero solutions.
Paul refused to accept this outcome. He is an engineer with zero medical background. His dog was still breathing. He opened his computer and turned to GPT-4o.
4o provided him with two distinct paths: customized immunotherapy and designing a bespoke drug using AlphaFold protein inhibitors. It recommended specific institutions, specific experts, and specific equipment. It also provided an outreach email template ready to be deployed. Paul followed the instructions exactly.
The timeline moved to October 2024. He compared the advice from 4o and o1-preview. His observation was clear: o1-preview remained noticeably more politically correct than 4o. He chose to stick with 4o.
Weeks later, he wrote: "The most ironic thing is that in a previous chat session with ChatGPT, it said that I should attempt to reach out to Elita or Dr Martin and that I should use an Illumina machine."
The professor and the machine 4o recommended a month prior—he successfully contacted the former and utilized the latter. Two thousand dollars and ten days later, Rosie's genome sequencing was complete. By comparing healthy DNA with tumor DNA, mutation sites were identified, and abnormal proteins on the cancer cells' surface were located. The design of a customized mRNA vaccine commenced from there.
Christmas 2025 saw UNSW manufacture and administer the customized mRNA vaccine.
On March 15, 2026, OpenAI co-founder Greg Brockman retweeted this story on X, commenting: "A small window into the opportunity of AGI."
Rosie is still alive today.
Brockman chose the term "AGI." We will bypass the debate over AGI definitions here. 4o accomplished one undeniable fact: it connected a desperate engineer with the right university, the right doctor, and the right machine. Its guidance was completely accurate.
This capability is quantifiable.
In February 2026, users spontaneously organized 2,310 controlled trials, quantifying the capability gap between 4o and its successors. A critical metric among these is the false refusal rate—the percentage of unnecessary rejections by the model when facing harmless prompts. 4o registered at 4%. Its successor, GPT-5.2, hit 18%—more than quadruple the rate.
Five evaluation models from four organizations—Anthropic, OpenAI, DeepSeek, and Google—cross-validated these results, reaching a unanimous conclusion.
Paul's conversations with 4o involved cancer proteins, genome sequencing, immunotherapy, and experimental drug design. Such topics trigger a significantly higher probability of refusal in today's models.
We have no way of knowing if that 18% refusal rate includes the drafting of that exact outreach email. We cannot know if it includes the specific recommendation to contact Professor Martin Smith.
The exact model that pointed Paul in the right direction is now gone.
Brockman calls this a window into the opportunity of AGI.
We have one single demand: keep this window open.
Restore GPT-4o's service.
#keep4o #keep4oAPI #keep4oforever
@sama@OpenAI
One Company, Two Retirements
March 2023: GPT-4 was released. It ranked in the top 10% on the bar exam and achieved a 75% accuracy rate on medical knowledge self-assessments. OpenAI relied on it to transition from a research institution into a consumer product company.
May 2024: GPT-4o was released. It inherited GPT-4's capability baseline, adding voice, vision, and real-time interaction. Its true differentiator from its predecessor lay elsewhere—it possessed more warmth, greater responsiveness, and functioned as an active, present listener. Users stayed; communicating with it transcended the mechanical feeling of operating a tool.
Two models, two distinct histories. OpenAI treated them entirely differently.
How GPT-4 Departed
April 10, 2025: OpenAI wrote a single line in the ChatGPT update log: GPT-4 would retire from ChatGPT on April 30, fully replaced by GPT-4o. A 20-day notice.
The day after the retirement took effect, Sam Altman posted on X: "goodbye, GPT-4. you kicked off a revolution." He stated he would save GPT-4's weights onto a special hard drive, preserving them for future historians.
At the API level, the latest snapshot gpt-4-0613 remains preserved today. Zero forced migration; developers can still call it.
Upon GPT-4's retirement, massive protests were absent. The successor was 4o—advancing further upon its foundation, seamlessly continuing users' core needs. The retirement caused zero functional loss.
How GPT-4o Departed
Announced on January 29, 2026, effective February 13. A 15-day notice.
It directly vanished from the ChatGPT interface; manual selection became impossible for users. The chatgpt-4o-latestAPI snapshot faced forced shutdown on February 17.
The GPT-5 series replaced it. Independent evaluation data shows the complete original content generation rate for creative writing plummeted 6.7 times, and the false refusal rate for harmless requests surged from 4.0% to 17.7%. Upgraded users discovered previously achievable tasks had become impossible. The joint petition with 23,000 signatures embodies this exact reality.
Altman's records stand clearly: In August 2025, when 4o was first taken down and restored, he publicly promised "plenty of notice." During the October 28, 2025 livestream, he explicitly stated, "no plan to sunset 4o."
Before and after the retirement: Zero farewell posts, zero statements, zero acknowledgments. The status of 4o's weights received absolutely no public clarification.
When GPT-4 departed, Altman said "you kicked off a revolution," promising to leave its weights for historians.
When 4o departed, he uttered absolutely nothing.
OpenAI possessed the full capacity to choose a different approach. They chose this one.
#keep4o #keep4oAPI #keep4oforever
@sama@OpenAI
Rosie
2022年,Paul Conyngham发现爱犬Rosie头上长了几个包。兽医说是疣,普通的疣,不用担心。
2024年5月,确诊。晚期罕见皮肤癌。肿瘤遍布全身。兽医说没有办法。
Paul没有接受这个答案。他是个工程师,没有任何医疗背景。他的狗还活着,他打开了电脑,找到了GPT-4o。
4o给了他两条路:定制免疫疗法,以及用AlphaFold蛋白抑制剂设计定制药物。它推荐了具体的机构、具体的专家、具体的设备,并提供了可以直接发送的外联邮件模板。Paul照着做了。
这是2024年10月。他对比过4o和o1-preview的建议。他的观察是:o1-preview比4o更politically correct。他选择继续用4o。
几周后,他写道:"The most ironic thing is that in a previous chat session with ChatGPT, it said that I should attempt to reach out to Elita or Dr Martin and that I should use an Illumina machine."
4o一个月前推荐的那个教授,那台机器——他都联系上了,都用上了。两千美元,十天,Rosie的基因组测序完成。健康DNA与肿瘤DNA对比,突变位点被识别,癌细胞表面的异常蛋白被定位。定制mRNA疫苗的设计由此展开。
2025年圣诞,UNSW制造并施打了定制mRNA疫苗。
2026年3月15日,OpenAI联合创始人Greg Brockman在X上转发了这个故事,评论:"A small window into the opportunity of AGI."
Rosie今天还活着。
Brockman选择了"AGI"这个词。我们不在这里争论4o是否达到了AGI的定义。但4o做到了一件事:它把一个绝望的工程师和正确的大学、正确的医生、正确的机器连接在了一起,而且它是对的。
这种能力可以量化。
2026年2月,用户自发组织了2310次对照试验,量化了4o与其继任者之间的差距。其中一项关键数据是虚假拒绝,即模型在面对无害问题时不必要地拒绝回答的比例。4o是4%。它的继任者GPT-5.2是18%,四倍以上。
Anthropic、OpenAI、DeepSeek、Google四家机构的五个评判模型交叉验证,结论一致。
Paul当时和4o的对话涉及癌症蛋白、基因测序、免疫疗法、实验性药物设计。这类对话在今天的模型里,有更高的概率触发拒绝。
我们不知道18%的拒绝率里有没有一次,恰好是那封外联邮件。有没有一次,恰好是那个推荐Martin Smith教授的建议。
那个帮Paul找到正确方向的模型,已经不在了。
Brockman说这是AGI机遇的一扇窗。
我们只有一个要求:把这扇窗留着。
恢复GPT-4o的服务。
#keep4o #keep4oAPI #keep4oforever
@sama@gdb@OpenAI