OpenAI Dots

Status: Rolling out on Pro and Business Premium since September 29, 2026. Beta on Enterprise, Edu, and Healthcare workspaces, off by default.

What a dot actually is

A normal ChatGPT conversation is reactive. You ask, it answers, and then nothing happens until you type again. A dot is OpenAI’s attempt at the opposite: an AI agent that holds onto a goal you gave it, keeps working in the background, and comes back to you when it has a result or needs a decision.

OpenAI announced Dots at DevDay on September 29, 2026. Each dot runs on GPT-6 Astra (OpenAI’s frontier model at launch, covered on the ChatGPT page) and gets its own cloud computer with a browser. You can open that computer to watch what the dot is doing or step in yourself. That separation matters: the work does not depend on your laptop staying awake, and the dot’s browsing happens in its own environment rather than yours.

You start with one “primary” dot, which you can name. OpenAI has also described “specialist” dots owned by an organization, each with its own identity, credentials, and tools, but those are limited to focused enterprise pilots for now.

How it differs from asking ChatGPT

The useful mental model is a colleague with a standing assignment rather than a search box. You might ask a dot to keep an eye on a competitor’s pricing page, collect customer feedback from a shared inbox each week, or prepare a status report every Friday from a project tracker. It schedules those runs itself (you manage them from a Scheduled tab in the dot’s profile) and reaches you where you already are: in ChatGPT on desktop, web, or mobile, or in Slack and Microsoft Teams. Context carries across those channels. Text messaging is in a limited beta for US Pro users.

Dots connect to apps through ChatGPT’s plugin ecosystem, which OpenAI puts at more than 4,000 apps. If you allow it through the ChatGPT desktop app, a dot can also work with local files and start Codex or ChatGPT Work tasks. That local access is off by default and can be revoked.

There are things it cannot do yet. At launch a dot cannot make phone calls and does not get its own email address, although it can use a personal email account you connect.

Memory, and why deleting it is awkward

A dot inherits your existing ChatGPT memories and forms its own, including from what it reads in connected apps. That is what makes it useful over weeks rather than one session, but it has a practical consequence worth knowing before you connect anything sensitive. According to OpenAI’s help documentation, the only way to delete a dot’s own memories is to delete the dot. Disconnecting an app stops future access but does not erase what the dot already learned from it. We cover that general pattern in Disconnecting AI Does Not Delete Data, and it applies here directly.

Resetting a dot deletes its conversations, memories, and scheduled tasks. Files, Codex threads, and ChatGPT conversations it created are stored separately and stay.

Approvals and safety controls

OpenAI built several layers around what a dot is allowed to do, and they are worth understanding rather than trusting blindly.

Custom Rules assign each kind of action one of four behaviors: act without asking, act only if pre-approved (meaning you explicitly asked for it in your prompt), ask first, or hand off to you. This is a form of authorization gate, and setting it conservatively at first is sensible.

Proactive research is the mode a dot uses when it is digging into something on its own initiative. In that mode it only has read-only tools: it cannot send messages, edit app content, or control the browser or computer.

Auto-review checks actions that affect accounts or share information against your instructions, rules, and OpenAI’s safety requirements. You cannot turn Auto-review off with Custom Rules, and some tasks, such as changing passwords, always stay with you. OpenAI also says it can pause or stop a dot if monitoring flags a safety concern.

For sign-ins, OpenAI says saved passwords are used without being exposed to the model. 1Password for Agents lists ChatGPT Work and Dots as a coming-soon integration for the same purpose.

None of this makes a dot infallible. OpenAI itself says dots can make mistakes and that consequential work should be reviewed. An agent reading your inbox and the open web is also exposed to indirect prompt injection, where instructions hidden in content try to steer it.

Privacy

Content from Business, Enterprise, and Edu workspaces is not used for model training by default. On personal plans, training is governed by your ChatGPT data controls. OpenAI says proactive research results and notes are not trained on directly, but information from them may be used depending on your settings, and human review can happen in limited cases such as safety reviews. If you are evaluating Dots for work, read our AI data retention explainer and check your workspace settings before connecting client data.

What it costs to run

Your first dot is included on eligible plans. OpenAI says conversations with a dot do not count toward ChatGPT usage limits, while deeper background work draws on a plan allowance and anything a dot runs in Codex or ChatGPT Work counts toward those limits normally. The first month was described as extended usage, and OpenAI has not published what happens after that. An always-on agent that checks things on a schedule can quietly consume more than you expect, so it helps to set a task budget in your head, if not in the product.

Where Dots fit well, and where they don’t

Dots make the most sense for ongoing, low-stakes-per-step work: monitoring a set of sources, keeping a running research brief, preparing recurring reports, chasing project updates, or doing administrative follow-ups that you review before anything goes out. The read-only research mode and approval rules fit that kind of work.

They fit less well where a single wrong action is expensive, such as moving money or sending messages to clients without review, and where you need full control over where data goes. Regional limits also matter: Pro access excludes the EEA, Switzerland, and the UK at launch.

Alternatives

• Claude Duet, Anthropic’s proactive agent, in research preview for US Pro and Max users
• Meta Muse, Meta’s personal agent with a free tier
• OpenClaw, a self-hosted, open-source personal agent if you want to run it on your own machine
• Daemons by Charlie Labs, always-on agents focused on engineering work
• Manus, a general agent for longer multi-step tasks

For background on the category, see proactive agents, approval policies, persistent agent state, long-running agents, and our guide to computer-use agents.

Compare: Gemini Tasks vs OpenAI Dots, OpenAI Dots vs Claude Duet, OpenAI Dots vs Meta Muse, OpenAI Dots vs OpenClaw, and the three-way Dots vs Duet vs Muse guide.