# OpenAI Will Eventually Build an OS
**作者**: CHOI
**日期**: 2026-04-27T17:03:47.000Z
**来源**: [https://x.com/arrakis_ai/status/2048810350821478545](https://x.com/arrakis_ai/status/2048810350821478545)
---
I think OpenAI will eventually build an operating system.
Codex is becoming increasingly popular, and OpenAI is expanding it into something far larger than a coding assistant. It is becoming a space where multiple agents can work at once, where real work can be delegated, and where documents, slides, spreadsheets, dashboards, browser workflows, images, and deployment all begin to converge.
On the surface, Codex can look like a super app for developers. But the more important question is not how many features OpenAI is adding to Codex. The real question is what kind of market OpenAI is preparing for inside Codex.
At first, Codex looked like a coding tool. Recent updates point in a different direction. Multiple agents can work in parallel. Files can be created and edited. Documents can be revised. Browsers can be inspected. Images can be generated. Apps can be deployed. It looks like a bundle of developer features, but underneath that, I see a deeper shift in how software is created, executed, and distributed.
Something similar is happening inside ChatGPT. Apps in ChatGPT already showed that users can call external apps directly inside a conversation, receive interactive app experiences, and discover or install apps through an app directory. With the Apps SDK, OpenAI gave developers a way to build app experiences that live inside ChatGPT. In other words, apps are no longer only destinations outside the browser. They are starting to enter the conversation itself.
That matters because the user's starting point is changing. Until now, the default flow was to open a browser, search, visit a website, install an app, log in, and then complete the task. In the next interface, the user may simply say, "Do this for me" inside ChatGPT. ChatGPT can call the right app, Codex can create a missing function, or an external app can render directly inside the conversation.
This does not mean the browser disappears. It means the browser may no longer be the front-facing canvas that directs the user's attention. It may become infrastructure in the background, used by agents to fetch information, render functionality, and connect services.
The meaning of the app store also changes. In the smartphone era, an app store was where finished apps were listed, discovered, downloaded, reviewed, and monetized. In the AI era, users may create the productivity tool they need in the moment, run it immediately, and distribute it if it proves useful. Buying a domain, choosing a server, configuring deployment, and waiting for app review could all become compressed or invisible.
When those steps shrink, people stop "visiting" a specific website or app just to use a function. They begin calling the function into their own conversation and workflow. This is where something like a Codex Store starts to make sense: an environment where users build apps with Codex, deploy them directly, and let others run them with visible permissions and safety boundaries.
This is why Codex's deployment direction matters. If Codex can deploy web apps to services like Cloudflare, Netlify, Render, and Vercel, that is not just a convenience feature. It is a sign that creation and distribution are being pulled into the same workspace. The user creates, tests, fixes, deploys, and updates software in one loop. At that point, Codex is no longer just a coding tool. It starts to feel like a software publishing environment.
Vibe coding is pushing this shift even faster. The difficulty of building websites, games, internal tools, and utilities is falling quickly. The next bottleneck is no longer just making the thing. It is deploying it, securing it, managing keys, handling payments, showing permissions, updating dependencies, and deciding whether the generated app is safe enough to run.
Google is attacking this bottleneck through its own strengths. Gemini, Google AI Studio, Firebase, and Firebase Studio point toward a world where a user can go from idea to app to deployment through Google's cloud infrastructure. That is a powerful strategy. Google starts from infrastructure, hosting, and distribution.
OpenAI starts from a different place. It starts from the conversation, the user's intention, the user's context, and the execution of agents. If ChatGPT manages context, Codex creates apps and automations, and the Apps SDK renders external apps inside ChatGPT, the traditional app store becomes less central. Users call apps from the flow of work. If the app does not exist, they create it. If it works, they distribute it.
This is where the market question becomes clear. If OpenAI remains only a company that provides powerful AI models, then it risks becoming a feature inside platforms controlled by Apple, Google, Microsoft, Samsung, browsers, and app stores. The next market will be decided by who owns the entry point, where user context accumulates, which apps are executed, and who controls payment, identity, permissions, and distribution.
From Sam Altman's point of view, this is a strategic fork. OpenAI can remain a model supplier, or it can try to own the next entrance to personal computing.
In the smartphone era, Apple and Google created the market by controlling the app store. Developers built apps. Users installed them. Apple and Google controlled distribution, payment, permissions, and trust. In the AI era, that structure can change. It is not enough to generate apps. The more AI-generated apps appear, the more important review, sandboxing, trust, permissions, key management, payment, and updates become.
That is why the software transition is also an operating system transition. Apps do not disappear. But the way apps are conceived, created, executed, distributed, reviewed, and permissioned changes. If OpenAI wants to own this market, a model alone is not enough. It needs the user interface, the execution environment, the distribution layer, identity, payment, security, memory, and eventually the device.
The meaning of an app changes as well. In the smartphone era, an app was a finished product. You searched for it, installed it, and operated within the functions it provided. In the AI era, an app can become more fluid. A user might say, "Create a dashboard that analyzes this report," "Turn this material into a presentation webpage," "Build a family travel budget calculator," or "Make this game idea playable as a web game." The app is generated and executed in context.
An app becomes less like a fixed product and more like an executable unit made from a user's intent, data, permissions, model, and interface. It can be created today, revised tomorrow, shared next week, and later converted into an automation.
This is also how I read Codex's image generation, editing, and GIF creation features. If a user can generate the assets they need inside a frontend or UI workflow without leaving the tool, the boundaries between planning, design, development, and deployment become weaker. PDF viewers, HTML viewers, embedded browsers, UI annotation, file preview, image generation, and deployment can look like small updates when viewed separately. Taken together, they pull the full software lifecycle into OpenAI's environment.
Computer Use is important in a different way. It is less about Codex today and more about what an OpenAI OS would need tomorrow. An AI OS cannot replace every existing app overnight. Banking apps, enterprise systems, shopping sites, hospital booking systems, accounting tools, and government services will continue to exist. Computer Use gives AI a way to operate existing software through the screen, mouse, and keyboard. In a future OpenAI OS, it could become the compatibility layer between the legacy software world and the agentic AI world.
This is why the hardware story matters. OpenAI has already formalized its work with Jony Ive, LoveFrom, and io, and the io Products team later joined OpenAI. Reports have also suggested that OpenAI is preparing its own smartphone to compete with the iPhone, with MediaTek, Qualcomm, Luxshare, and a possible 2028 timeline being discussed. Whether or not those reports become the final product roadmap, the strategic logic is clear. If OpenAI wants to own the user interface, the sensors, and the most personal data layer, it eventually has to confront hardware and OS.
An AI phone is not interesting because it runs ChatGPT well. It becomes interesting when the user's location, voice, gaze, routines, work habits, conversation history, and content preferences accumulate inside one system, and agents use that context to call apps, make payments, book services, and recommend actions.
Even ordering morning coffee could become a different kind of interaction. The user would not open a cafe app, choose a branch, and manually pay. The AI could consider the user's first meeting location, commute route, sleep quality, and caffeine sensitivity, then place the order at the right time. At that point, apps are no longer destinations the user opens. They become functional pieces the OS calls and assembles in the background.
Recommendation also changes. YouTube, Netflix, music, news, documents, calendars, and notes are currently recommended inside separate platforms. A personal AI OS could recommend across the user's full context. It could suggest the video you need now, the explanation that helps you understand a document, the automation that saves your week, or the learning path that fits your habits.
This is the age of the hyper-personalized OS. It is also the age of the hyper-biased OS. A system that hears everything, remembers everything, and adapts everything to you is extremely convenient. It can also trap you inside a narrower worldview. Today's social media algorithms bias users inside individual platforms. A personal OS can shape search, content, work, consumption, learning, and relationships at once.
As AI-generated content explodes, people will become more tired of information overload and overstimulating content. In a world flooded with generated material, people care less about what can be made and more about what can be trusted. The next platform advantage may come from trust: provenance, copyright, labeling, review, transparency, and recommendation standards.
This may also raise the value of offline experience. If digital spaces become crowded with AI content and social media trust declines, people may place more value on physical gatherings, offline entertainment, experiential activities, and real communities. The future AI OS may not compete only by increasing screen time. It may compete by helping users navigate and enhance real-world experiences.
Identity becomes central in this future. If OpenAI directly controls the OS and the device, World ID becomes strategically relevant in a different way. When AI agents can pay, install apps, send messages, submit documents, and agree to contracts on behalf of a user, the system must prove that the action came from a real person and that the person delegated that authority.
Passwords and SMS codes are not enough for that kind of delegated world. A proof-of-personhood system such as World ID, which combines iris-based human verification with zero-knowledge proofs, could become part of the future identity stack. World ID may look separate today, but if OpenAI builds its own device and OS, the value changes. Biometric verification, agent delegation, payment approval, token usage, and app execution permissions could all become connected inside the device.
In an agent-driven world, the question is not simply whether an account is logged in. The deeper question is what a real human has authorized an AI to do. Iris-based verification has serious privacy and regulatory risks, but the strategic point remains. The value of an AI phone may not be camera resolution or screen size. It may be the ability to bind context, identity, payment, token usage, and agent permissions into one trusted system.
Pricing may also change. If an AI-native phone emerges, the center of the phone bill may move from data to tokens. Today's smartphone plans are built around network access and data usage. An AI-native phone's core cost may come from inference, agent execution, memory retrieval, tool use, instant app generation, image generation, and video generation.
Users may end up paying less for gigabytes and more for token usage, action credits, agent runs, and context storage. In a more distant future, if satellite or space-based networks make physical connectivity nearly free, the cost users feel may shift from connection to intelligence. The phone bill becomes a token bill. The monthly plan depends on how deeply the user runs a personal AI OS, how much context they keep, how many agents they run in parallel, and how many apps they create and execute.
This is why I think OpenAI will eventually build an OS. Not because Codex is growing quickly. Not because ChatGPT has more features. The market itself is pushing in that direction. If AI remains a feature inside someone else's app, it remains part of someone else's platform. If AI is going to complete the user's work directly, it has to become the platform.
Codex is where this shift first appears in software work. Computer Use is where legacy software gets absorbed. Hardware is where OpenAI tries to own the user interface and the personal data layer. Together, these point toward a new operating environment.
It may not be called iOS or Android. It may begin as a workspace inside ChatGPT, a development environment inside Codex, or an execution environment that combines Computer Use with the Apps SDK. But when users delegate intent, AI creates apps, operates existing apps, deploys results, manages permissions, handles payments, remembers context, and runs across a personal device, the system has moved beyond the category of an app.
At that point, it is an operating system.
## 相关链接
- [CHOI](https://x.com/arrakis_ai)
- [@arrakis_ai](https://x.com/arrakis_ai)
- [4.5K](https://x.com/arrakis_ai/status/2048810350821478545/analytics)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [1:03 AM · Apr 28, 2026](https://x.com/arrakis_ai/status/2048810350821478545)
- [4,591 Views](https://x.com/arrakis_ai/status/2048810350821478545/analytics)
- [View quotes](https://x.com/arrakis_ai/status/2048810350821478545/quotes)
---
*导出时间: 2026/4/28 20:44:23*
---
## 中文翻译
# OpenAI 终将打造一个操作系统
**作者**： CHOI
**日期**： 2026-04-27T17:03:47.000Z
**来源**： [https://x.com/arrakis_ai/status/2048810350821478545](https://x.com/arrakis_ai/status/2048810350821478545)
---
我认为 OpenAI 最终会打造一个操作系统。
Codex 正变得日益流行，而 OpenAI 正在把它扩展成一个远超编程助手的存在。它正在成为一个多个智能体可以同时工作、真实工作可以被委派的空间，文档、幻灯片、电子表格、仪表盘、浏览器工作流、图像和部署都开始在这里汇聚。
表面上，Codex 看起来像是一个面向开发者的超级应用。但更重要的问题不是 OpenAI 正在向 Codex 添加多少功能。真正的问题是 OpenAI 正在 Codex 内部为一个怎样的市场做准备。
起初，Codex 看起来像是一个编程工具。但近期的更新指向了另一个方向。多个智能体可以并行工作。可以创建和编辑文件。可以修改文档。可以检查浏览器。可以生成图像。可以部署应用。这看起来像是一堆开发者功能的集合，但在其之下，我看到的是软件的创建、执行和分发方式正在发生更深刻的转变。
类似的事情也正在 ChatGPT 内部发生。ChatGPT 中的应用已经表明，用户可以在对话中直接调用外部应用、接收交互式应用体验，并通过应用目录发现或安装应用。借助 Apps SDK，OpenAI 为开发者提供了一种构建可驻留在 ChatGPT 内部的应用体验的方式。换句话说，应用不再只是浏览器之外的目的地。它们开始进入对话本身。
这一点很重要，因为用户的起点正在改变。迄今为止，默认流程是打开浏览器、搜索、访问网站、安装应用、登录，然后完成任务。在下一个界面中，用户可能只需在 ChatGPT 内部说一句“帮我做这个”。ChatGPT 可以调用合适的应用，Codex 可以创建缺失的功能，或者外部应用可以直接在对话中渲染。
这并不意味着浏览器会消失。它意味着浏览器可能不再是引导用户注意力的前台画布。它可能变成后台的基础设施，供智能体用来获取信息、渲染功能和连接服务。
应用商店的意义也在改变。在智能手机时代，应用商店是成品应用被上架、发现、下载、评价和变现的地方。在 AI 时代，用户可以即时创建他们所需的生产力工具，立即运行，并在证明有用后进行分发。购买域名、选择服务器、配置部署以及等待应用审核，都可能被压缩甚至变得不可见。
当这些步骤缩减时，人们就不再为了使用某个功能而专门“访问”特定的网站或应用。他们开始把功能调用到自己的对话和工作流中。这正是 Codex Store 这类事物开始变得合理的地方：一个用户用 Codex 构建应用、直接部署它们，并让其他人在可见的权限和安全边界内运行它们的环境。
这就是为什么 Codex 的部署方向很重要。如果 Codex 能够将 Web 应用部署到 Cloudflare、Netlify、Render 和 Vercel 等服务上，那就不只是一个便利功能。这是一个信号，表明创建和分发正在被拉入同一个工作空间。用户在一个循环中创建、测试、修复、部署和更新软件。到那时，Codex 就不再只是一个编程工具。它开始像一个软件发布环境。
氛围编程（Vibe coding）正在进一步加速这一转变。构建网站、游戏、内部工具和实用程序的难度正在迅速下降。下一个瓶颈不再仅仅是把东西做出来，而是部署它、保障其安全、管理密钥、处理支付、展示权限、更新依赖，以及判断生成的应用是否足够安全可以运行。
Google 正凭借自身优势攻克这一瓶颈。Gemini、Google AI Studio、Firebase 和 Firebase Studio 指向这样一个世界：用户可以通过 Google 的云基础设施，从想法到应用再到部署一气呵成。这是一个强大的战略。Google 从基础设施、托管和分发起步。
OpenAI 则从不同的地方起步。它从对话、用户意图、用户上下文以及智能体的执行起步。如果 ChatGPT 管理上下文，Codex 创建应用和自动化，Apps SDK 在 ChatGPT 内部渲染外部应用，那么传统应用商店的重要性就会下降。用户从工作流程中调用应用。如果应用不存在，他们就创建它。如果它好用，他们就分发它。
正是在这里，市场问题变得清晰起来。如果 OpenAI 仅仅停留在一家提供强大 AI 模型的公司，那么它就有可能沦为苹果、Google、微软、三星、浏览器和应用商店所控制的平台中的一个功能。下一个市场将由谁来决定：谁拥有入口，用户上下文在哪里积累，哪些应用被执行，以及谁控制支付、身份、权限和分发。
从 Sam Altman 的角度来看，这是一个战略岔路口。OpenAI 可以继续做模型供应商，也可以尝试掌控个人计算的下一个入口。
在智能手机时代，苹果和 Google 通过控制应用商店创造了市场。开发者构建应用。用户安装应用。苹果和 Google 控制着分发、支付、权限和信任。在 AI 时代，这一结构可能改变。仅仅生成应用是不够的。AI 生成的应用出现得越多，审核、沙箱、信任、权限、密钥管理、支付和更新就变得越重要。
这就是为什么软件转型同时也是操作系统的转型。应用不会消失。但应用的构思、创建、执行、分发、审核和授权方式都将发生改变。如果 OpenAI 想要拥有这个市场，仅靠模型是不够的。它需要用户界面、执行环境、分发层、身份认证、支付、安全、记忆，最终还需要设备。
应用的含义也在发生变化。在智能手机时代，应用是一个 finished 产品：你搜索它、安装它，并在它提供的功能范围内操作。在 AI 时代，应用可以变得更加流动。用户可能会说：“创建一个分析这份报告的仪表板”、“把这份材料做成一个演示网页”、“做一个家庭旅行预算计算器”，或者“把这个游戏创意做成可玩的网页游戏”。应用是在具体情境中生成并执行的。
应用不再像一个固定的产品，而更像一个由用户意图、数据、权限、模型和界面组成的可执行单元。它可以今天创建，明天修改，下周分享，之后再转化为一个自动化流程。
这也是我解读 Codex 的图像生成、编辑和 GIF 创建功能的方式。如果用户可以在前端或 UI 工作流内生成所需的素材而无需离开工具，那么规划、设计、开发和部署之间的界限就会变得更模糊。PDF 查看器、HTML 查看器、内嵌浏览器、UI 标注、文件预览、图像生成和部署，单独看可能像小更新。但放在一起，它们就把完整的软件生命周期拉进了 OpenAI 的环境。
Computer Use 的重要性则以另一种方式体现。它与今天的 Codex 关系较小，更多关系到明天的 OpenAI 操作系统需要什么。AI 操作系统不可能一夜之间取代所有现有应用。银行应用、企业系统、购物网站、医院预约系统、会计工具和政务服务仍将继续存在。Computer Use 让 AI 有办法通过屏幕、鼠标和键盘操作现有软件。在未来的 OpenAI 操作系统中，它可能成为旧软件世界与智能体 AI 世界之间的兼容层。
这就是为什么硬件的故事很重要。OpenAI 已经正式确定了与 Jony Ive、LoveFrom 和 io 的合作，io Products 团队后来也加入了 OpenAI。还有报道称，OpenAI 正准备推出自己的智能手机与 iPhone 竞争，联发科（MediaTek）、高通（Qualcomm）、立讯精密（Luxshare）以及可能的 2028 年时间表都在讨论之列。无论这些报道是否会成为最终的产品路线图，其战略逻辑是清晰的。如果 OpenAI 想要掌握用户界面、传感器和最私人的数据层，它最终必须直面硬件和操作系统。
AI 手机的有趣之处不在于它能很好地运行 ChatGPT。它变得有趣，是因为用户的位置、语音、视线、日常习惯、工作模式、对话历史和内容偏好都积累在同一个系统中，而智能体利用这些上下文来调用应用、完成支付、预订服务和推荐行动。
甚至早上点咖啡也可能变成一种不同类型的交互。用户不需要打开咖啡店应用、选择门店、手动支付。AI 可以综合考虑用户第一个会议的地点、通勤路线、睡眠质量和咖啡因敏感度，然后在合适的时间下单。到那时，应用不再是用户打开的目的地，而是操作系统在后台调用和组装的功能组件。
推荐也会改变。YouTube、Netflix、音乐、新闻、文档、日历和笔记目前是在各自独立的平台内进行推荐的。而个人 AI 操作系统可以基于用户的完整上下文进行推荐。它可以推荐你现在需要的视频、帮助你理解某个文档的讲解、为你节省一周时间的自动化流程，或者适合你习惯的学习路径。
这是超个性化操作系统的时代，也是超偏见操作系统的时代。一个倾听一切、记住一切并为你调整一切的系统极其便利，但也可能把你困在一个更狭窄的世界观里。今天的社交媒体算法在单个平台内使产生偏见。而个人操作系统可以同时影响搜索、内容、工作、消费、学习和人际关系。
随着 AI 生成内容的爆发，人们会对信息过载和过度刺激的内容感到更加疲惫。在一个生成内容泛滥的世界里，人们关心的不再是“能制造什么”，而是“什么可以信任”。下一个平台优势可能来自信任：来源可溯、版权、标注、审核、透明度和推荐标准。
这也可能提升线下体验的价值。如果数字空间充斥着 AI 内容、社交媒体信任度下降，人们可能会更加重视线下聚会、线下娱乐、体验式活动和真实的社区。未来的 AI 操作系统可能不再仅靠增加屏幕时间来竞争，而是通过帮助用户更好地驾驭和丰富现实世界的体验来竞争。
身份认证在这一未来中变得至关重要。如果 OpenAI 直接掌控操作系统和设备，World ID 将以另一种方式具有战略意义。当 AI 智能体可以代表用户进行支付、安装应用、发送消息、提交文件和签署合同时，系统必须证明该操作来自一个真实的人，并且这个人授予了相应权限。
对于这样一个委托代理的世界，密码和短信验证码是不够的。像 World ID 这样的个人身份证明系统，将基于虹膜的人类验证与零知识证明相结合，可能成为未来身份认证体系的一部分。World ID 今天看起来似乎是独立的存在，但如果 OpenAI 构建自己的设备和操作系统，其价值就会改变。生物识别验证、智能体授权、支付批准、代币使用和应用执行权限，都可能全部关联到设备内部。
在一个由智能体驱动的世界里，问题不再仅仅是账户是否已登录。更深层次的问题是，一个真实的人类授权了 AI 去做什么。基于虹膜的验证存在严重的隐私和监管风险，但其战略要点依然成立。AI 手机的价值可能不在于摄像头分辨率或屏幕尺寸，而在于将上下文、身份、支付、令牌使用量和智能体权限绑定到一个可信系统中的能力。
定价模式也可能发生变化。如果 AI 原生手机出现，手机账单的核心可能会从流量转向令牌。如今的智能手机套餐是围绕网络接入和数据使用量构建的，而 AI 原生手机的核心成本可能来自推理、智能体执行、记忆检索、工具使用、即时应用生成、图像生成和视频生成。
用户最终可能会为千兆字节付更少的钱，而为令牌使用量、操作积分、智能体运行次数和上下文存储支付更多。在更遥远的未来，如果卫星或天基网络使物理连接几乎免费，用户感受到的成本可能会从连接转向智能。手机账单变成令牌账单。月度套餐取决于用户运行个人 AI 操作系统的深度、保留多少上下文、并行运行多少个智能体，以及创建和执行多少个应用。
这就是为什么我认为 OpenAI 最终会构建一个操作系统。不是因为 Codex 增长迅速，也不是因为 ChatGPT 功能更多，而是市场本身正在朝这个方向推进。如果 AI 仍然只是别人应用中的一个功能，它就仍然是别人平台的一部分。如果 AI 要直接完成用户的工作，它就必须成为平台本身。
Codex 是这种转变在软件工作中首次显现的地方。Computer Use 是传统软件被吸收的地方。硬件则是 OpenAI 试图掌控用户界面和个人数据层的地方。这三者共同指向一个新的操作环境。
它可能不会被称为 iOS 或 Android。它可能最初是 ChatGPT 内的一个工作区、Codex 内的一个开发环境，或者一个将 Computer Use 与 Apps SDK 结合的执行环境。但当用户委托意图、AI 创建应用、操作现有应用、部署结果、管理权限、处理支付、记住上下文，并在个人设备上运行时，这个系统就已经超越了应用的范畴。
到那时，它就是一个操作系统。
## 相关链接
- [CHOI](https://x.com/arrakis_ai)
- [@arrakis_ai](https://x.com/arrakis_ai)
- [4.5K](https://x.com/arrakis_ai/status/2048810350821478545/analytics)
- [升级至 Premium](https://x.com/i/premium_sign_up)
- [2026年4月28日 凌晨1:03](https://x.com/arrakis_ai/status/2048810350821478545)
- [4,591 次浏览](https://x.com/arrakis_ai/status/2048810350821478545/analytics)
- [查看引用](https://x.com/arrakis_ai/status/2048810350821478545/quotes)
---
*导出时间: 2026/4/28 20:44:23*
