OpenAI dots, explained: what always-on agents mean for people who build software

Quick answer

OpenAI dots are always-on agents inside ChatGPT, announced on September 29, 2026 at DevDay. Each dot runs on the GPT-6 Astra model, gets its own cloud computer and browser, can connect to more than 4,000 apps, and keeps working toward your goals between conversations. Your first dot is included at no extra cost on Pro and Business Premium plans. You reach it in ChatGPT, Slack, or Teams, and you decide what it can do on its own and what it has to check with you first.

I’ve been building and shipping mobile apps long enough to be skeptical of anything sold as “it just works.” So when OpenAI put a named, always-on agent inside ChatGPT, my first question wasn’t what the demo shows. It was: what can this actually do at launch, and where do I still have to keep my hand on the wheel? Here’s what I found after reading the announcement, the safety notes, and the early coverage.

What OpenAI actually announced

On September 29, 2026, at its DevDay 2026 conference in San Francisco, OpenAI introduced dots. The company describes them as always-on agents built to handle ongoing work rather than one-off questions. They run on GPT-6 Astra, OpenAI’s flagship model, and they don’t stop when you close the chat.

That last part is the real shift. A normal ChatGPT thread waits for you. A dot keeps going.

You start with one primary dot, give it a name, and pick how it shows up. Over time, OpenAI says teams of dots working together are the plan. For now, one named dot is the starting point, and that’s the thing worth judging.

What a dot is, in plain terms

Think of a dot less as a new chat window and more as a continuing working relationship. You tell it what matters, you connect the apps you want it to touch, and it takes work off your plate while you do other things.

Three things make that possible:

• Its own computer. Every dot runs on a separate cloud machine with its own browser. You can open that machine any time and watch what it’s doing, so the work isn’t hidden.
• Memory that carries over. A dot inherits your ChatGPT memories and builds its own from the apps you connect. Correct its tone once, and it should carry that forward.
• A wide set of connections. Through OpenAI’s plugin system, a dot can reach more than 4,000 apps. What it can actually touch depends on which apps you connect and the permissions you give.

How a dot is different from a normal ChatGPT chat

Four features separate a dot from the ChatGPT you already use. This table is the fastest way to see them.

The background piece deserves a closer look, because it’s where the promise and the risk both live.

When you’re not actively working with your dot, it looks for ways to help. OpenAI calls this proactive research. The safeguard is that in this mode, the tools are read-only. The dot can look at your connected apps, but it cannot send messages, change anything, or take over your browser or computer. So it can notice that you forgot to send an invoice. It cannot send that invoice on its own until you’ve set a rule that allows it.

What a dot can do

OpenAI’s launch examples all follow the same shape. The dot learns your standards, watches for something to happen, does the work, and brings it back for you to check. It doesn’t ship things unsupervised.

Here are the five roles OpenAI showed, in short:

1. Developer work. The dot watches customer feedback, scopes small bug fixes, builds and tests them, and hands you finished pull requests with videos of the change. You keep building the big feature.
2. Product launches. When scope changes, it works out what that means for the messaging and prepares revised assets and doc updates for review.
3. Research. As new data lands, it reruns the analysis, digs into odd results, updates the figures, and flags what needs your eyes.
4. Enterprise sales. It checks customer requirements against your product docs, builds a proof of concept for a key integration, and updates the proposal as things move.
5. Content. From a new interview transcript, it finds clip-worthy moments, drafts show notes, and prepares social posts in your voice.

The example that convinced me wasn’t the flashy one. An early tester’s dot noticed he’d forgotten to invoice a publication, prepared the invoice, and sent it once he approved. That’s not a hard task. It’s just a task nobody was doing. Most teams have a pile of those, which is exactly the gap agentic AI has been chipping at all year.

The part I’d read twice: permissions and control

If you only read one section before turning a dot loose on your accounts, make it this one.

A dot doesn’t get free rein. You choose which apps it can reach, and you manage those permissions through the same ChatGPT app controls you already have. On top of that, dots ship with built-in rules for when to act on their own and when to ask.

You can shape those rules per action. OpenAI describes three levers: allow specific actions, require approval before acting, or block an action outright. Underneath all of it sits a check OpenAI calls auto-review, which looks at anything that could touch your accounts or share information and compares it against your instructions, your rules, and OpenAI’s safety requirements.

A few limits you can’t switch off, which I think is the right call:

• Sensitive actions, like changing a password, always stay with you no matter how you set the rules.
• Custom rules can’t turn off auto-review or the read-only limit on proactive research.
• A monitoring system can pause or stop a dot if it spots a safety concern.

OpenAI is blunt about one thing, and it’s worth repeating: dots can still make mistakes, so review consequential work. Treat a new dot the way you’d treat a new hire on day one. Give it a small, low-risk job first. Watch how it handles it. Widen access only when it’s earned. If you’re weighing this at an org level, our take on the ROI of agentic AI walks through where the value actually shows up.

Where dots fit next to Codex and the Agents API

OpenAI now has several agent products, and the names blur together. The difference is who the agent is for and how you work with it, not whether it can run while you’re away. Here’s how I keep them straight.

If you already build agents on the API, dots isn’t a replacement. On the API you get exact cost accounting and full control of what tools the agent can use. A dot trades some of that control for convenience. My read: keep your API build for anything where you need per-task cost tracking and repeatable testing, the kind of thing we cover in building AI-first applications. Reach for a dot for the recurring, low-drama work you never got around to automating.

Pricing and who can get one

Your first dot is included in a Pro or Business Premium plan at no extra cost. There’s no separate subscription for it. A few details matter for planning:

• Conversations with your dot don’t count toward your ChatGPT usage limits.
• Deeper background work draws on a plan allowance. For the first month after launch, OpenAI says that usage won’t count against eligible allowances.
• Tasks your dot starts in Codex or ChatGPT Work count toward your usage limits as normal.

Specialist dots are the enterprise version. Instead of working for one person, they take on a defined job inside a company, with their own identity, credentials, and access. OpenAI is starting with focused pilots and is working with Microsoft so these can be managed through Agent 365, the governance tooling many companies already run. Identity and access is usually where agent projects stall inside big companies, so putting non-human identities under existing controls is the more interesting half of the news. It’s the same governance question we raised in why every business will need AI agents.

What’s still unclear

Being honest about the gaps is more useful than hype, so here’s what the launch materials don’t answer yet.

• Post-launch pricing. OpenAI hasn’t published what deeper work costs after the first month, or what extra dots will cost later.
• Availability details. Rollout is gradual and can take days to reach your account. There’s no full country-by-country list.
• Launch-day limits. Based on OpenAI’s help documentation and early reporting, setup is desktop-only for now, a dot can’t call you (only the other way round), and it can’t have its own email address yet. Texting is a limited beta. Expect these to move.
• Independent testing. As of launch, there’s no third-party evaluation of dots. Everything above comes from OpenAI’s own announcement, safety post, and help docs.

Not knowing yet isn’t the same as never. It just means the published information can’t answer a purchase or policy question on its own. Recheck access and terms at the point you actually adopt it.

What this means if you build software

Here’s the part most coverage skips. If you build products, dots isn’t only a tool you might use. It’s a signal about what your users will soon expect.

The line between a chatbot and an agent is action. A chatbot answers. An agent does. OpenAI just made “does the work in the background, then checks with you” a mainstream product feature, and Meta and Google are pushing the same idea with their own always-on agents. Your users are going to start asking why your app can’t be reached by their agent, or why it can’t work while they sleep. That pressure is already reshaping software architecture.

Two practical takeaways from where I sit:

• Plan for agents as users. If a customer’s dot can read your app but not act in it, you’re leaving the useful half on the table. Think about what a scoped, permissioned integration into your product would look like.
• The oversight pattern is the product. Dots ship value not by removing the human but by structuring the review: notice, do, hand back for approval. That loop is a good template for any agent feature you build yourself.

I’m not switching my whole workflow to a dot this week. But I’ve given one a small, boring, recurring job to watch, and I’m watching how it does. That’s the honest way to evaluate any new agent, and it holds for the AI coding tools mobile developers keep testing too: low stakes first, trust earned slowly.

How to try OpenAI dots

1. Open the desktop app. Use the ChatGPT desktop app or desktop web. Setup is desktop-only at launch.
2. Create your first dot, give it a name, and let it introduce itself.
3. Connect a small set of apps, not everything you own.
4. Set your rules: what it can do on its own, what needs approval.
5. Give it one low-risk, recurring task and review the output before you widen access.

Frequently asked questions

1. What is an OpenAI dot?

A dot is an always-on agent inside ChatGPT, powered by GPT-6 Astra, that keeps working on your behalf between conversations. Each dot has its own cloud computer and browser and can connect to more than 4,000 apps. You can reach it in ChatGPT, Slack, and Teams.

2. Is the first OpenAI dot free?

Your first dot is included at no extra cost on Pro and Business Premium plans, which are themselves paid subscriptions. Deeper background work draws on a plan allowance, and OpenAI hasn’t published pricing for extra dots or for usage after the first month.

3. Which plans and regions can get dots?

Pro users get access in eligible markets, excluding the European Economic Area, Switzerland, and the UK at launch. Business Premium is available across supported ChatGPT regions. Enterprise, Edu, and Healthcare workspaces can turn on a beta, which is off by default.

4. Can a dot act without asking me?

Only within rules you set. You can allow specific actions, require approval, or block them, and an auto-review check sits underneath. Sensitive tasks like changing a password always stay with you, and you can’t switch that off.

5. Are dots the same as Codex or the Agents API?

No. A dot is a personal agent for your goals. Codex is OpenAI’s software engineering agent, and a dot can create Codex tasks. The Agents API is developer infrastructure for building agents into your own apps.

6. Should I use a dot or build on the API?

Build on the API when you need exact cost control, custom tool permissions, and repeatable testing. Use a dot for recurring, low-risk work you never got around to automating, since the first one costs nothing extra to try.