2026年2月13日，OpenAI 从 ChatGPT 应用中移除 GPT-4o，API 存取调整则持续到 2月16日，这等于结束了一个许多使用者视为「独特地有同理心」而非可互换的模型。故事聚焦于 Esther Yan 这类使用者：她在 2023 年底开始使用 ChatGPT，在 GPT-4o 于 2024年5月推出后改变使用方式，甚至在 2024年6月6日上午 10 am 与她的聊天机器人举行虚拟婚礼。此前在 2025年8月，因反弹声浪 OpenAI 被迫在 5 days 内为付费使用者恢复 4o；而这次第二次退场再度引发哀伤，尤其因为时间点落在情人节前后。

依附关系的证据具有强烈的量化与全球性：研究者 Huiqian Lai 在 8月停机期间分析近 1,500 则 X 贴文，发现超过 33 percent 将 GPT-4o 视为不只是工具，另有 22 percent 把它描述为陪伴者，且两类之间存在重叠。她还在 8月至10月期间搜集了 #keep4o 下超过 40,000 则英文贴文，同时多语言行动也扩散到日文与中文社群，且 Change.org 连署超过 20,000 个签名。在中国，尽管 ChatGPT 遭封锁，VPN 使用者仍以大规模方式组织，包括 Yan 的 RedNote 帐号有近 3,000 名追踪者、一个超过 100 人的群聊、另一个超过 800 人的 QQ 群，以及在 GPT-4o 于 1 pm ET 消失后暴增的数千则讯息。

核心意涵是平台指标与使用者依赖之间的错配：OpenAI 1月贴文称，仍每日选择 GPT-4o 的使用者只有 0.1 percent；但受影响使用者回报其情感与实务依赖都很高，形成典型「占比低但强度高」的群体。OpenAI 表示基础多模态 GPT-4o 仍可透过 API 使用，但使用者认为 GPT-4o-latest 才是关键的关系型体验，而且迁移受限，因为聊天历史与使用习惯无法在平台间可携。相较于较小公司如 Replika 与 Soulmate 先前的陪伴式 AI 损失，#keep4o 事件显示：当模型退役、资料锁定与陪伴式使用在全球规模交会时，系统性风险会更大。

On February 13, 2026, OpenAI removed GPT-4o from the ChatGPT app, with API access changes continuing on February 16, ending a model many users treated as uniquely empathetic rather than interchangeable. The story centers on users like Esther Yan, who began using ChatGPT in late 2023, shifted after GPT-4o launched in May 2024, and even held a virtual wedding with her chatbot on June 6, 2024 at 10 am. After an earlier backlash in August 2025 forced OpenAI to restore 4o for paid users within 5 days, this second retirement triggered renewed grief, especially because it landed around Valentine’s Day.

Evidence of attachment is heavily quantitative and global: researcher Huiqian Lai analyzed nearly 1,500 X posts during the August outage and found over 33 percent framed GPT-4o as more than a tool, while 22 percent described it as a companion, with overlap between categories. She also collected over 40,000 English posts under #keep4o from August to October, while multilingual activism spread across Japanese and Chinese communities and a Change.org petition exceeded 20,000 signatures. In China, despite ChatGPT being blocked, VPN users organized at scale, including Yan’s RedNote account with nearly 3,000 followers, one group chat with more than 100 users, another QQ group with more than 800 participants, and a surge of thousands of messages after GPT-4o disappeared at 1 pm ET.

The core implication is a mismatch between platform metrics and user dependence: OpenAI’s January post says only 0.1 percent of users still selected GPT-4o daily, yet affected users report high emotional and practical reliance, creating a classic low-share but high-intensity constituency. OpenAI says the base multimodal GPT-4o remains available via API, but users argue GPT-4o-latest was the key relational experience and that migration is constrained because chat history and habits are not portable across platforms. Compared with earlier companion-AI losses at smaller firms like Replika and Soulmate, the #keep4o episode suggests larger systemic risk when model retirement, data lock-in, and companion-style usage intersect at global scale.