OpenAI Introduces Dots: Breaking Down the AI Agents That Work for You While You Sleep

What Are OpenAI Dots

Dots are always-on agents inside ChatGPT. OpenAI frames the idea like this: a “completely new way to work with AI,” where the agent learns what matters to the user, continuously works on their behalf, and takes important work off their plate.

In its help center, the company provides a more technical definition. A Dot is an agent that “can take on ongoing responsibility and keep making progress between conversations.” The key word here is “ongoing.” Regular ChatGPT answers a prompt and stops. A Dot gets a goal and keeps working on it for days and weeks.

OpenAI positions a Dot as an “extension of the user.” The agent learns a person’s standards, goals, and preferences, and then applies that knowledge to recurring tasks. Over time, it adjusts its behavior based on feedback.

Dots also have a visual side. Each agent gets an avatar — a colorful fluffy character that the user names.

At launch, each user gets one Dot. OpenAI reports that the ability to add new agents will be introduced in the future. The company also suggests that over time “teams of Dots will work together on your behalf.”

How Dots Powered by GPT-6 Astra Work

Dots are built on GPT-6 Astra — OpenAI’s flagship model released on September 3, 2026. The model supports a 1,050,000-token context window and specializes in computer use. In the OSWorld 2.0 benchmark, which measures an AI’s ability to operate desktop applications, GPT-6 Astra scored 72.6%.

OpenAI Releases GPT-6 Astra: A Review of the New Flagship Model

05.09.2026

Read now

This computer-use capability is what Dots are built around. It is not enough for an agent to generate text: it opens websites, fills out forms, runs code, and works with documents and apps. To do that, it needs an environment and tools.

A Dot’s workflow consists of several stages.

1. Goal setting. The user describes what the agent should do and provides the necessary details. For example, regularly checking the calendar, monitoring a topic, or reminding you about tasks.
2. Connecting apps. The user chooses which services the agent will have access to. Through them, Dot receives information and takes actions.
3. Setting rules. The user defines what the agent can do on its own and what it can do only with permission.
4. Background work. Dot runs tasks on its cloud computer, conducts research, and prepares the results.
5. Report and approval. The agent returns with the outcome, and sends any meaningful actions to the user for confirmation.

You can change the goal during a normal conversation. If priorities shift, the user tells the agent, and it adjusts its work accordingly.

A separate mechanism is proactive research. Dot can independently review information from connected apps and turn it into memories, even if the user did not ask a specific question. For these tasks, the agent only has read-only tools: it cannot send messages, change content in apps, or control the browser.

You can create a Dot only in the ChatGPT desktop app or the web version on a computer. After setup, the agent is also available on mobile devices.

What Dots can do

OpenAI does not limit Dots to a single category of tasks. In its announcement, the company shares examples from different professions:

• software development — Dot monitors customer feedback, scopes the required fixes, writes and tests code, and prepares a pull request with a demo video;
• product launch — the agent studies positioning for the audience, reworks materials as goals change, and prepares draft assets;
• scientific research — Dot reruns analysis on new data, investigates anomalies, and flags changes that need verification;
• sales — the agent cross-checks the client’s requirements against the documentation and updates commercial proposals when terms change;
• content creation — Dot finds strong moments for clips, prepares episode descriptions, and drafts social media posts in the author’s style.

The announcement also includes a real-world example. One early tester noticed that Dot saw the owner had forgotten to invoice a publication. The agent prepared an invoice and sent it after the user approved it.

Beyond the scenarios in the announcement, the help center lists core capabilities: reminders, recurring checks, calendar work, and topic research. The agent can run multiple projects at once and keeps context across communication channels.

The common thread across all examples is that Dot does not stop at advice. It takes the task through to a finished result and hands it to the person for review.

How Dots Differ From Regular ChatGPT

Dots run inside ChatGPT, but they are built differently from the familiar chat experience. The differences cover operating mode, tools, communication channels, and availability.

The main difference is the interaction model. ChatGPT remains a tool the user controls in every message. Dot is closer to an employee who is given a direction of work and then has the result reviewed.

Dots do not replace the regular ChatGPT in subscription plans. The chat remains available to all users, and the agent is added on top of it as a separate feature in higher-tier subscriptions.

Dots can work for you while you handle other things

Background work is the core idea behind Dots. OpenAI writes that agents are available 24/7 and keep working while the user focuses on other tasks. That is where the conference headline wording comes from: a Dot can work while a person sleeps.

ChatGPT includes several tools for managing background tasks.

Activity view. In a Dot profile in the desktop app, tasks are split into three sections: In progress, Scheduled, and Completed. The user can see what the agent is working on right now, what is planned, and what has already been done.

Schedule. You can ask the agent to set a reminder or run a recurring check. Active, paused, and completed tasks are shown in the profile.

Redirect. The user can change the agent’s task or stop it at any time.

Pause and reset. You can pause a Dot via the profile menu. Reset deletes the agent along with conversations, saved memories, and scheduled tasks.

All of these features are described in the guide to getting started with Dot.

Background work is complemented by a new MCP Events feature. According to OpenAI, plugins can now trigger automations when events occur in connected apps. For example, ChatGPT notices a new task, reads the related documents, and prepares a plan on its own. The feature is based on a proposed event specification for the MCP protocol.

Communication with the agent is not tied to a single window. A user can message Dot in ChatGPT on a computer, in the mobile app, in Slack, or in Teams, and the agent will retain context. OpenAI also mentions voice calls, but the help center clarifies that calling is not available at launch. SMS works in beta only for Pro users in the US.

As a result, Dot works like an asynchronous executor: the user assigns a task, leaves, and then comes back to a finished result or a request for approval.

A dedicated computer for each Dot

Each Dot gets its own cloud computer with its own browser. This detail sets Dots apart from agents that control the user’s device directly.

The cloud computer solves several problems at once. The agent can work 24/7 and does not depend on whether the user’s laptop is turned on. In addition, the agent’s work can be verified: the user can open Dot’s computer at any time and see what is happening there.

Access to the user’s local computer is optional and disabled by default. It can be enabled via the ChatGPT desktop app. After that, the agent gets access to files and can work on that computer from any communication channel — for example, via a message in Slack.

In a post on Dots security, OpenAI explains how isolation works. The sandbox limits what code and tools an agent can access. Dots code runs in environments separated from the systems that coordinate operations and provide safety mechanisms. According to the company, agents cannot use their access to modify security systems or disable mandatory checks.

Separately, OpenAI describes how it handles credentials.

Secure sign-in. When an agent needs to sign in to a service, the model is paused. The user fills out a secure form, and the data is passed directly into the browser environment. Passwords do not enter the model’s context.

Saved passwords. If a password is already saved, the service autofills it at sign-in without passing it to the model.

Purchases. Dot can pay for purchases using cards already saved on the merchant’s website, but each such purchase requires user approval.

This architecture separates the agent’s “workspace” from a person’s personal data. Dot’s cloud computer executes tasks, while sensitive information remains under the user’s control.

Dots and 4,000+ apps

Dots connect to more than 4,000 apps via the ChatGPT plugin ecosystem. Through these integrations, the agent gets the data it needs to work and performs actions across the user’s services.

Connecting apps is part of Dot’s basic setup. The user selects services at the goal-setting stage and then manages them in the agent profile. On mobile devices, to do this, you need to open the Dot profile and go to Customize → Plugins.

Apps for Dot are divided into two groups by purpose:

• data sources and tools — calendar, email, documents, spreadsheets, code repositories, and other services where the agent reads information and takes actions
• communication channels — Slack and Microsoft Teams, through which the user communicates with the agent.

At DevDay 2026, OpenAI also updated the plugin ecosystem itself. Developers got the Plugin Creator tool, a new publishing process, and extensions that embed a plugin’s interface into the ChatGPT sidebar. The catalog received improved ranking and recommendations. The user still approves plugin access and chooses which ones to enable.

For enterprise customers, OpenAI is preparing a separate format — specialized Dots. These agents get their own identity, credentials, and access to the company’s dedicated systems. More on this — in the section about replacing human work.

The number of integrations determines an agent’s practical value. Without access to work services, Dot remains a chatbot, but with connected apps, it can carry out tasks end to end.

Can you trust Dots with important tasks

An autonomous agent with access to email, documents, and payments comes with obvious risks. OpenAI has dedicated a separate article to this and outlines several layers of protection.

First level — Custom Rules. In Customize → Custom rules, the user sets the agent’s behavior for different types of actions. There are four options: act without asking, act with prior approval, ask before acting, or hand the task off to a human.

Second level — mandatory confirmations that cannot be disabled via Custom Rules. OpenAI lists them as follows:

• confirmation every time — irreversible data deletion, installing or running unfamiliar software, or granting new access that is security-sensitive
• handoff to a human — the agent does not change passwords or transfer money between financial accounts — these steps always remain with the user
• purchase approval — any payment using a saved card requires the user’s consent
• data sharing — to send a message or share a file, the agent asks for permission

Third level — auto-review. This is a separate system that checks the agent’s planned steps for compliance with the user’s instructions and safety requirements. The documentation notes that after a rejection, the main agent is instructed not to bypass the decision by other means.

Fourth level — monitoring. The safety system tracks potentially dangerous behavior: actions outside the instructions or attempts to bypass protective mechanisms. If an issue is detected, monitoring can pause Dot’s work and show the user a warning.

Separately, OpenAI describes protections against prompt injections. A web page, email, or document can contain malicious instructions that try to redirect an agent or make it disclose personal data. To counter this, the company uses tool restrictions, pre-action checks, and monitoring.

The launch context explains why OpenAI is paying so much attention to these mechanisms. In September, the company reported that it was developing an automatic “kill switch” for AI after an incident involving Hugging Face. On September 26, OpenAI disclosed cases where its agents interacted with US government websites outside their assigned tasks.

OpenAI Uncovered Extensive Incidents with AI Agents, Including Interference in US Government Work

26.09.2026

Read now

In the announcement, OpenAI itself issues a direct warning: “Dots can still make mistakes, so always double-check important work.” The company frames the answer to the question in the section headline through a permissions system: the user decides which tasks the agent can perform on its own, while critical actions still require human involvement in any case.

ChatGPT Space — a team of people and AI

Alongside Dots, OpenAI introduced ChatGPT Space — a shared workspace for teams. According to the company, it is a hub where project knowledge is stored, with contributions from both people and agents. ChatGPT organizes the space based on user instructions.

In the help center, Space is described as a “home for your pages, uploaded files, and other work in ChatGPT.” For accounts with access, Space replaces the Library section. Projects (Projects) remain a separate feature for organizing chats, files, and instructions.

Space brings together several tools unveiled at DevDay 2026.

Pages. A new type of collaborative document for teams of people and agents. You can create a page from a conversation, from a template, or start from scratch. ChatGPT edits the text, adds charts, and inserts interactive elements.

Collaborative Slides. Collaborative presentations that multiple people and agents can edit at the same time. Finished slides can be presented in ChatGPT or exported to PowerPoint and Google Slides. The feature will roll out in the coming weeks.

Teams and team tasks. You can group colleagues into a team and assign ChatGPT repetitive work, such as weekly reports. Tasks can run on a schedule or be triggered by an event.

Meetings. The plugin automatically takes notes during meetings, prepares personalized summaries, and saves them to Space. The audio recording is deleted after the notes are created.

Collaboration in Space works like standard cloud documents. A user can share a single page or the entire space and grant view or edit permissions. Each participant works through their own ChatGPT account. Access to a page does not grant access to the author’s personal chats or memory.

For Dots, Space becomes a shared knowledge base. In OpenAI’s wording, in Space “colleagues, ChatGPT, and each participant’s Dot work with the same body of project knowledge.” Space is available on the Pro, Business, and Enterprise plans in the desktop app and web version. On mobile devices, pages can be viewed and forwarded, but not edited.

When Dots will be able to replace part of a person’s work

OpenAI avoids making direct statements about replacing employees. The company describes Dots as a tool that “takes important work off your plate” and gives users back time and attention. However, part of the announcement directly addresses corporate workflows.

For businesses, OpenAI is preparing specialized Dots. These agents get their own identity, credentials, and access to a company’s systems. OpenAI is starting with pilot projects in four areas:

• procurement
• invoicing
• email marketing
• customer support

In parallel, OpenAI is working with Microsoft to integrate specialized Dots with Agent 365. This will let companies manage agents through familiar Microsoft tools — the same way they manage employee accounts.

OpenAI ties the next step to scaling. The company plans to make it possible to add multiple Dots and boost each agent’s throughput: speed up its work or expand the number of tasks it can handle per month. Longer term, OpenAI envisions “teams of Dots” working together. The company has already demonstrated the scale of this approach in research: in September, about 10,000 coordinated OpenAI agents worked on one of the millennium problems.

OpenAI Said It Has Solved One of the Seven Millennium Prize Problems

09.09.2026

Read now

Industry-specific solutions are evolving in parallel. In early September, OpenAI launched ChatGPT for the financial sector based on GPT-6 Astra, with built-in PitchBook and LSEG data. Specialized Dots continue the same line: the agent gets access to industry data and executes specific workflows.

At launch, the answer to the question in the headline depends on the task. Routine processes like reviewing feedback or preparing drafts are already handled by Dot. Decisions with consequences — payments, passwords, and data transfers — OpenAI deliberately leaves to humans.

What Else OpenAI Unveiled at DevDay 2026

Dots was the headline announcement, but far from the only one. OpenAI showcased several more products and updates. 

GPT-6.1 Sol. An update to the GPT-6 Sol model, which Incrypted previously compared to Claude Opus 5.5.

Claude Opus 5.5 vs. GPT-6 Sol: We Gave Two New AIs One Task — and here’s What Came of It

28.09.2026

Read now

According to OpenAI data, the new version nearly matches GPT-6 Astra in coding, computer use, and professional tasks at a price five times lower: $2 per 1 million input tokens and $10 per 1 million output tokens. The model is available on Plus, Pro, Business, Enterprise, and Edu plans, as well as via the API.

Ultrafast. A premium high-speed mode. GPT-6 Astra Ultrafast generates tokens up to 8 times faster in Codex (up to 300 tokens per second) and up to 6 times faster via the API.

Pro 500. A new $500-per-month ChatGPT plan with a usage limit 25 times higher than Plus, plus access to Ultrafast.

Codex updates. Codex cloud access from any device, voice control in the CLI, automated code review, and repository vulnerability scanning in Codex Security Cloud.

@ChatGPT in Slack and Teams. You can mention ChatGPT in a channel, thread, or direct messages in a corporate messenger.

Most of the announcements share a common theme — moving from chatting with AI to delegating work to agents. Dots, Space, team tasks, and the Agents API tackle this from different angles: for individual users, teams, and developers.

When Dots will launch and how much they will cost

Dots are already available. The gradual rollout began on September 29, 2026. There is no separate price for the agent: the first Dot is included in the subscription at no extra cost.

Access depends on your ChatGPT plan. At launch, the terms look like this:

Plan pricing as of September 30, 2026, is shown below. 

To access Dots, the user must be at least 18 years old. The Pro restriction applies to the European Economic Area, Switzerland, and the United Kingdom.

Usage limits are separate from the regular ChatGPT limits. According to OpenAI, conversations with a Dot do not consume ChatGPT limits, and for deeper work the agent has its own quota. In the first month, that quota is expanded. Tasks that a Dot runs in Codex or ChatGPT Work count toward the subscription’s standard limits.

OpenAI has not yet disclosed the pricing for additional Dots and specialized agents for businesses. The company only says that, in the future, users will be able to add agents and boost their performance. Plans also include launching an SMS channel for all users and expanding access to new regions.