What Is a "Dot"? OpenAI's Always-On Agent, Explained

For three years, "using ChatGPT" meant opening a tab, typing something, and waiting for an answer. OpenAI's newest product breaks that loop on purpose. A dot, announced at OpenAI's DevDay on September 29, 2026, is an agent that keeps running after you stop looking at it — monitoring a project, using software on your behalf, reacting to new information, and bringing finished work back for your approval when it's ready, not when you happen to check in.

That distinction — an agent that initiates rather than one that only responds — is the entire premise of the product, and it's worth sitting with before getting into features. A dot isn't a smarter chatbot. It's closer to a very literal-minded employee who never logs off, reachable through ChatGPT, Codex, Slack, and Microsoft Teams, with phone and an audio interface described as "coming soon."

What a dot actually has access to

Every dot gets its own cloud computer and its own browser — infrastructure that is explicitly separate from your machine unless you connect it. That separation matters more than it sounds: it means a dot's default working environment is a sandbox, not your laptop, and any bridge to your actual desktop is something you grant rather than something it assumes.

From that sandbox, a dot reaches the outside world through OpenAI's plugin ecosystem, which OpenAI says now spans more than 4,000 applications — the same connective tissue that already threads through ChatGPT's existing tool integrations. In practice, that's the difference between "an AI that can talk about your CRM" and "an AI that can open your CRM, filter a view, and export a list," provided the plugin exists and you've authorized it.

OpenAI's own framing leans on a handful of concrete jobs: a developer's dot watching customer feedback channels, drafting fixes, and returning pull requests with demo videos attached; a product team's dot revising launch materials automatically as the underlying product changes; a sales dot checking a prospect's stated requirements against internal documentation to assemble a proof-of-concept; a content dot turning a single meeting transcript into a week's worth of clips, notes, and social posts. None of these are hypothetical showcase demos — they're the specific scenarios OpenAI chose to lead with, which tells you where the company expects the first real usage to land. We go deeper on the developer case in The Developer's Dot.

Who gets one, and when

Access on day one is narrower than the marketing suggests. The first dot is free and included for ChatGPT Pro and Business Premium subscribers in eligible markets — PYMNTS reports that Pro users in the EEA, Switzerland, and the UK are excluded from this initial rollout, while Business Premium access extends across all currently supported ChatGPT regions. Usage in the first month reportedly won't count against normal plan allowances, though what happens to pricing after that is, as of this writing, undisclosed. Enterprise, Edu, and Healthcare customers get a slower, admin-gated beta rather than self-serve access — a rollout pattern we look at more closely in Enterprise, Edu, and Healthcare: The Slow Lane for Dots.

Additional dots beyond the free first one, and any dedicated "specialist" dots built for narrower jobs, are both things OpenAI has said will exist without saying what they'll cost. That's a genuinely undisclosed gap, not a rounding error, and we've dedicated an entire article to pricing the unpriced rather than guess here.

The permission layer, briefly

OpenAI's answer to "how do you let something act autonomously without it acting recklessly" comes in three pieces, each covered in more depth later in this series: Custom Rules, which let you pre-approve, block, or require sign-off on specific categories of action; Auto-Review, a background check that runs a dot's more consequential moves against your instructions, your rules, and OpenAI's own safety baseline before anything executes; and an Activity View where you can watch what a dot has been doing and step in mid-task. Sensitive operations — changing a password is the example OpenAI uses — are reserved for a human, full stop, regardless of how the rules are configured.

A dot's most interesting capability isn't what it can do while you're watching. It's what it's allowed to do — read-only "proactive research" in the background — while you aren't.

That background capability, which OpenAI calls proactive research, is explicitly scoped to read-only tools: a dot can look, but in that idle mode it cannot send a message, edit a file, or take control of a browser session. It's a narrow permission, deliberately so, and it's the mechanism that lets a dot notice something has changed without also being the mechanism that lets it act unsupervised on what it noticed.

Where this sits next to everything else

Dots arrived in the same nine-day stretch as xAI's Grok Bot expansion and Meta's Muse, three companies converging on the same idea — an agent with its own computer that outlives the chat window — from three different trust models. Dots vs Grok Bot vs Muse lays out how those permission philosophies actually diverge, which matters more than any feature checklist once you're deciding what to let one of these things touch.

It's also worth being precise about what a dot is not, because the category is genuinely new and easy to conflate with adjacent things. A dot is not a Claude Code Skill — a skill is a specific, inspectable, git-versioned capability you install into an agent you already run and already trust, not a standalone always-on coworker with its own cloud computer and its own login to your tools. They solve different problems at different layers, and a dot showing up on your team doesn't retire the case for the narrower kind of tooling aiskill.market catalogs. We take that comparison seriously, rather than dismissively, in Will Dots Replace Claude Code Skills, or Just Sit Beside Them?

For now, the honest summary is this: OpenAI has shipped a genuinely new interaction model — agents that act on a schedule and a set of standing permissions rather than a single prompt — with real, if partial, guardrails and a rollout that's deliberately staged rather than universal. What GPT-6 Astra, the model actually driving a dot's judgment, brings to that picture is where this series goes next.

Part 1 of 30 in the ChatGPT Dots series. Next: Inside GPT-6 Astra: The Model Powering Dots. Browse more builder insights or explore AI skills at aiskill.market.