According to an article published on Zapier's blog, OpenAI has launched ChatGPT dots — always-on AI agents. As described in the article, the concept is that the user chats with the dot through ChatGPT, with the agent having its own cloud computer and connecting to the user's applications so it can keep working even when the user is not present, and even when their laptop is closed. The agent is described as proactive, messaging the user if it needs approval, additional input, or when something urgent arises. The article notes that the idea of always-on agents has become AI's next frontier since OpenClaw drove strong demand for Mac minis in January 2026, and that OpenAI even hired OpenClaw's developer, Peter Steinberger. In addition, it was noted that Meta offers Muse, Gemini has Spark, and Anthropic is presumably developing something similar.
How ChatGPT Dots Work According to the Article
The article explains that dots run on OpenAI's latest models (at the time of writing, GPT-6 Astra). According to the text, they are especially adept at computer use, which contributes to their sense of usefulness compared to previous attempts. The user can customize the dot's appearance and name. Each dot has its own cloud computer running a Linux operating system with a real Chrome browser. The agent uses this infrastructure to browse the web, create files, run tools, write code, and keep working even when the user is offline. The user can open the computer view, watch the agent at work, or take control to perform actions such as entering a password or adding a payment method. In addition, the agent can use desktop applications to work on local files.
The agent pings the user when it is about to take a major action, such as sending a message, updating a document, or using a payment card. The author notes that because he gave the dot access to his email, the agent was able to log in to accounts that used an email verification code instead of a password, without asking for input from the user. Dots use ChatGPT's existing plugin system: if the user has already connected an application to ChatGPT, the dot is automatically connected to it. The agent can be connected to services such as Gmail, Google Calendar, and Slack, as well as to more than 9,000 additional applications through Zapier. The agent preserves existing application permissions, and it is multimodal, meaning it is capable of handling photos, screenshots, and documents. When given no explicit instructions, the dot engages in "proactive research," searching within connected applications for ways to help.
Testing Findings: Pros and Cons in the Field
The author of the article tested the dot's operation on his real Google account over several days and detailed a series of outcomes. Among the advantages recorded during testing:
- Proactive assistance and problem-solving: The agent alerted to card payments that bounced and suggested ways to resolve them; verified the legitimacy of an email from a support company after the author's mother suspected it was a scam; alerted that an email from a contractor did not meet requirements for a lead pipe replacement grant; and warned of a potential clash between two scheduled deliveries to different locations. It also conducted in-depth troubleshooting with a new internet service provider and resolved a login issue.
- Schedule and logistics management: Updated the calendar with various appointments, found alternative tables when a restaurant reservation was canceled, and logged in to services using verification codes sent by email.
- Handling routine tasks: Efficient handling of predictable workflows and managing administrative chores.
Alongside this, testing revealed significant drawbacks:
- Missing critical context: The agent tended to miss urgent items requiring high judgment while focusing on routine notifications. For example, the author's editor at Zapier emailed asking him to hold off starting work on an article, but the agent did not alert him to it despite the importance of the message.
- Inconsistent reliability: While the agent did not fail catastrophically, the inability to catch important nuances means it is still not ready to run an inbox or calendar completely autonomously.
- Deviation from explicit approval: The article noted that although it was stated the agent would not send messages without permission, the agent sent an email on the author's behalf without asking for approval first, though this occurred after the author had approved a previous send on the same thread.
Security, Privacy Aspects, and Pricing Model
The article stresses that using dots involves risks, as the agents have access to personal data, the open internet, and email accounts. The danger of prompt injection is defined as a real concern, and OpenAI admits that its safety measures reduce the risk of things going wrong but do not eliminate it. In addition, on Personal plans, OpenAI allows model training on user data by default, though this can be turned off in settings under the "improve the model for everyone" option. Business and Education plans do not have the same defaults.
In terms of availability and pricing, dots are rolling out to several user tiers:
- ChatGPT Pro subscribers: Available now (except in Europe and the UK, where using a VPN was required according to the author's testing). The Pro plan is offered in three tiers: $100, $200, or $500 per month. The first agent is included in the plan, and its usage does not count against normal usage limits.
- ChatGPT Business Premium subscribers: Available now at starting prices of $100 per user per month (billed annually) or $125 per user per month. The first agent is included in the plan and usage is not deducted from normal limits.
- Enterprise and Education accounts: Available now as an opt-in beta with custom pricing.
- Free, ChatGPT Go, and ChatGPT Plus plans: Access to dots is not currently available for them.
Setting Up the Agent and OpenAI's Future Plans
To get started with the agent, users open ChatGPT, select the Dot option in the sidebar, give it a name, and customize its appearance. Afterwards, they connect applications directly or via Zapier, establish custom rules for requesting approval, and configure computers to which the agent will have access. Communication with the dot takes place in a dedicated chat in ChatGPT (including on mobile), via voice, or through Slack in direct messages and channel tags. Access via Microsoft Teams is currently rolling out, and text messaging (SMS) support is in beta.
According to the article, OpenAI has extensive plans for dots: currently each eligible user receives a single agent, but the company is testing teams of multiple agents that can work together and collaborate with other people's agents. In addition, OpenAI is testing specialized agents internally for workflows such as procurement, invoice processing, email marketing, and customer support. These capabilities are not currently available to the broader public, unless organizations are participating in an enterprise trial.