OpenAI的新版Dots旨在独立处理工作任务，例如修复Slack中报告的错误，或在有人提出请求之前发送被遗忘的发票。与此同时，OpenAI还发布了价格更低的GPT-6.1 Sol模型。由于安全问题，高端型号GPT-6.1 Astra暂时搁置。

OpenAI在DevDay 2026开发者大会上推出了Dots。据该公司介绍，这些是全天候运行的智能体，能够接管持续进行中的任务，监控项目进展，并在需要用户关注时主动联系用户。每个Dot都拥有自己的云端计算机和浏览器，因此可以全天候工作。它可以在那里进行研究、分析数据、撰写文档，或通过Codex构建软件。Dots基于GPT-6 Astra运行。

从技术角度看，Dots与三周前推出的Meta的Muse智能体类似。据Meta称，Muse也在其受保护的独立计算机上运行，配备浏览器，在后台持续工作，并在执行敏感操作前寻求批准。

插件功能让Dot能够连接超过4,000款应用，并同步推进多个项目。OpenAI表示，该智能体会从反馈中学习用户的工作方式以及对良好结果的定义。用户可以通过ChatGPT中的聊天或语音通话与他们的Dot互动，也可以在Slack和Microsoft Teams中使用。OpenAI称，Dot在所有这些渠道中都能保持上下文连贯性。

其理念是，与Dot协作应更像是在与同事发消息交流。每个Dot都有名字和卡通头像。在OpenAI的演示中，桌面应用左侧显示聊天窗口，右侧第二个窗口则展示智能体正在处理的内容，无论是Google Slides演示文稿，还是带有浏览器、终端和文件存储的云端计算机。如果Dot处于控制状态，用户可以点击“接管”介入操作。

这与ChatGPT Work中的Computer Use功能存在主要区别：在后者中，用户需逐个启动任务，且智能体直接操作用户的屏幕。而Dot默认在其自己的计算机上工作，并在需要时将任务移交Codex处理。只有在获得许可的情况下，它才能访问用户的笔记本电脑。

OpenAI表示，在一个内部案例中，一个Dot追踪了Slack中报告的错误，直至生成可用的新版本构建。对于早期测试者而言，该智能体发现了一张被遗忘的发票，准备好后在测试者批准后发送了出去。每位用户在发布时将获得一个Dot，但OpenAI的长期计划是让整个团队的Dots协同工作。

OpenAI的演示还包括日常琐事，例如通过短信订购食物。该公司表示该功能即将推出。

Dots主动出击，但仅限读取权限

当没有人主动与之交互时，Dot会在后台寻找提供帮助的方式。OpenAI称之为“主动研究”。在此模式下，智能体仅使用只读工具，无法发送消息、更改内容或控制浏览器和计算机。

对于活跃任务，一套基本规则决定了Dot何时独立行动，何时先进行确认。自定义规则允许用户允许特定操作、要求对此类操作进行批准，或直接禁止它们。OpenAI表示，任何涉及账户或共享信息的操作在通过前，都会经过自动化检查以审查其是否符合这些规则。更改密码等敏感任务始终由人类掌控。

Dot还可以使用保存的密码，而模型本身永远不会看到这些密码。安全措施旨在阻止恶意指令，如果存在安全隐患，监控系统可以暂停或停止工作。活动视图（Activity View）会显示Dot正在处理的内容，包括在后台运行的任务。OpenAI承认Dots可能会犯错，并建议对任何具有实际后果的结果进行双重检查。

OpenAI表示，默认情况下，它不会使用来自Business、Enterprise和Edu工作区的内容进行训练。个人计划的用户可自行决定是否将其Dot的对话和工作内容用于训练。OpenAI还表示，它不会直接利用后台研究或Dot为自己保留的笔记进行训练。

欧洲用户还需等待

欧洲将迎来受限版发布。目前，Pro 订阅者只能在欧洲经济区、瑞士和英国以外的地区使用 Dots。Business Premium 客户可在所有支持地区使用，而 Enterprise（企业）、Edu（教育）和 Healthcare（医疗）客户则可在管理员开启后参与测试版体验。用户需先在桌面应用或浏览器中设置 Dot，随后即可在移动端使用。

首个 Dot 随订阅免费提供，与其聊天不计入使用量限制。它在 Codex 或 ChatGPT Work 中启动的任务则按常规计算用量。首月使用限额将予以扩大。未来如需更多 Dots、更高速度或更大工作量，则需额外付费。

OpenAI 正为企业推出承担组织内固定角色的专用 Dots。这些专用 Dot 拥有独立的身份和凭证，若企业有意，还可由 IT 部门提供计算机支持。OpenAI 表示，已在采购、发票处理、电子邮件营销、客户支持和合同管理等内部场景中对其进行了测试。此次推广将从各公司的试点项目开始。OpenAI 还与微软合作，将专用 Dots 引入 Agent 365——即微软的管理平台。

没有炒作成分的 AI 新闻 —— 由人工精选

订阅 THE DECODER，享受无广告阅读、每周 AI 通讯、每半年六次的独家“AI Radar”前沿报告、完整档案访问权限以及评论区互动功能。

立即订阅

OpenAI's new Dots are built to handle work tasks on their own, like fixing a bug someone reported in Slack or sending out a forgotten invoice, sometimes before anyone asks. Alongside them, OpenAI is releasing GPT-6.1 Sol, a cheaper model. The top-end GPT-6.1 Astra is being held back over safety problems.

OpenAI introduced Dots at its DevDay 2026 developer conference. According to the company, they're always-on agents that take over ongoing tasks, keep an eye on projects, and reach out when something needs the user's attention. Each Dot has its own cloud computer with a browser, so it can work around the clock. There it does research, analyzes data, writes documents, or builds software through Codex. Dots run on GPT-6 Astra.

Technically, Dots are similar to Meta's Muse agent, which launched three weeks earlier. According to Meta, Muse also runs on its own secured computer with a browser, keeps working in the background, and asks for approval before taking sensitive actions.

Plugins let a Dot connect to more than 4,000 apps and move several projects forward at the same time. OpenAI says the agent learns from feedback how its user works and what they consider a good result. Users can reach their Dot through chat or voice calls in ChatGPT, and also in Slack and Microsoft Teams. The Dot keeps its context across all of these channels, OpenAI says.

The idea is that working with a Dot should feel more like messaging a coworker. Every Dot has a name and a cartoon avatar. In OpenAI's demos, the desktop app shows the chat on the left. A second window on the right shows what the agent is working on, whether that's a Google Slides presentation or its cloud computer with a browser, terminal, and file storage. If the Dot is in control, users can click "Take over" to step in.

That's the main difference from Computer Use in ChatGPT Work, where users start each task individually and the agent operates their own screen. A Dot works on its own computer by default and hands tasks off to Codex when it needs to. It can only access the user's laptop with permission.

In one internal case, OpenAI says a Dot followed a bug reported in Slack all the way to a working new build. For an early tester, the agent spotted a forgotten invoice, prepared it, and sent it once the tester approved. Each user gets one Dot at launch, but OpenAI's long-term plan is whole teams of Dots working together.

OpenAI's demos also include everyday errands, like ordering food by text message. The company says that feature is coming soon.

Dots take initiative, but only with read access

When nobody is actively working with it, a Dot looks for ways to help in the background. OpenAI calls this "proactive research." In this mode, the agent only uses read-only tools, which can't send messages, change content, or control a browser or computer.

For active tasks, a set of ground rules decides when a Dot acts alone and when it checks in first. Custom Rules let users allow individual actions, require approval for them, or ban them outright. OpenAI says an automated check reviews any action that touches accounts or shares information against those rules before it goes through. Sensitive tasks like changing a password always stay with the human.

A Dot can also use saved passwords without the model ever seeing them. Safeguards are supposed to block malicious instructions, and a monitoring system can pause or stop the work if safety concerns come up. An Activity View shows what the Dot is working on, including tasks running in the background. OpenAI admits Dots can make mistakes and recommends double-checking any results that carry real consequences.

OpenAI says it doesn't train on content from Business, Enterprise, and Edu workspaces by default. Users on personal plans decide for themselves whether their Dot's conversations and work can be used for training. OpenAI also says it doesn't train directly on the background research or the notes a Dot keeps for itself.

European subscribers have to wait

Europe gets a restricted launch. For now, Pro subscribers can only use Dots outside the European Economic Area, Switzerland, and the UK. Business Premium customers get them in all supported regions, while Enterprise, Edu, and Healthcare customers can try a beta once their admins switch it on. Users set up their Dot in the desktop app or browser first, and after that it's also available on mobile.

The first Dot comes with the subscription, and chatting with it doesn't count against usage limits. Tasks it kicks off in Codex or ChatGPT Work count as usual. Limits are expanded for the first month. More Dots, extra speed, or more work volume will cost extra later on.

OpenAI is going further for businesses with specialized Dots that take on fixed roles inside an organization. These get their own identity, their own credentials, and, if the company wants, computers supplied by IT. OpenAI says it has tested them internally in purchasing, invoice processing, email marketing, customer support, and contract management. The rollout starts with pilot projects at individual companies. OpenAI is also working with Microsoft to bring the specialized Dots to Agent 365, Microsoft's management platform.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now