ChatGPT stops being a request and response loop
OpenAI dots, announced at DevDay on September 29, 2026, change the shape of the product rather than its speed. Until now ChatGPT has been a loop: you type, it answers, the thread goes quiet until you type again. A dot is a standing assignment. You hand it a name, a goal and a set of boundaries, and it keeps working toward that goal while you are asleep or stuck in a meeting.
Three things arrive on day one. Every dot gets its own cloud computer, with a browser and a filesystem that are not yours. It gets its own identity, separate from your account and from the plugins you have connected. Each user starts with one primary dot, which tells you how OpenAI is thinking about density for now. Pro and Business Premium subscribers in eligible markets get in first, and Enterprise, Edu and Healthcare workspaces need an admin to switch the beta on. There is no self-hosted path and no open weights. The model running underneath is GPT-6 Astra.
The honest framing is that most of this was already reachable. Codex tasks plus a connected plugin plus a browser automation library plus a cron job gives you a persistent agent that reads a shared inbox and files a report every Monday. I have built versions of that on my own boxes, and they failed in boring ways. A session cookie expired after nine days. A CSS selector changed and the scraper started returning empty strings. A retry fired twice and the same invoice reminder went out to a paying customer. None of that was hard to write. All of it was annoying to maintain.
What OpenAI is selling with dots is packaging and identity. One agent with its own machine, and a control layer a site owner without a DevOps background can reason about. The capability underneath is not the headline here. According to MarkTechPost’s writeup of the launch (opens in new tab), each dot runs on a separate cloud computer you can open and inspect at any time, which is the part that actually matters when something goes wrong at 3am and you need to know what it did.
Background
The trajectory matters more than the launch date. For two years, the way you made ChatGPT do ongoing work was to leave a tab open and re-prompt it, or to write a Codex task and fire it from a scheduler. Plugins multiplied because each one was a narrow bridge to a single app, and those connectors accumulated faster than anyone’s ability to audit them. Running a real deployment meant a folder of shell scripts, a token rotation calendar, and a quiet suspicion that the automation was doing more than you had asked. OpenAI eventually counted more than 4,000 plugin apps in that pile, which is a lot of credentials sitting in a lot of places.
Then the model underneath changed. In OpenAI’s latency simulations, GPT-6 Astra scores 72.6% on OSWorld 2.0 at roughly 40 minutes per task. GPT-5.6 Sol scored 65.7% at roughly 75 minutes. Read that twice: a higher score at nearly half the wall-clock time. Long-horizon agent work is bounded by how long a model stays coherent before it drifts, so pulling a task from 75 minutes down to 40 changes which jobs are worth delegating at all. Work that used to time out or wander off is now inside the budget.
Worth being precise about what those numbers are. They come from OpenAI’s own latency simulations, not from an independent use, so I would treat the direction as informative and the exact figures as provisional until someone reproduces them on their own tasks.
The competitive frame landed before dots did. Meta Muse arrived September 8, 2026, running on Muse Spark inside a Muse Secure VM, which is a dedicated cloud computer by another name. Its pull was the Meta app graph rather than the agent architecture itself. Instinct has stayed invite-only since August 2026, working out of your own computer and phone number, with no mobile app and no public pricing listed. So by the time OpenAI stepped up on September 29, the market already had one consumer-scale agent and one invitation-only experiment, and very little sitting in the middle for developers who wanted an agent with an actual machine behind it.
That gap is where dots are pointed.
What’s happening now: OpenAI dots at DevDay 2026
Dots shipped as a managed product and nothing else. There is no self-hosted build, no open-weights release, no Docker image you can drop on a VPS and point at your own storage. Pro and Business Premium subscribers in eligible markets are the first through the door. Enterprise, Edu and Healthcare workspaces sit behind an admin toggle that has to be flipped before anyone in the tenant sees the feature at all. Per MarkTechPost’s writeup of the DevDay announcement (opens in new tab), each user starts with one primary dot.
For anyone who has run an agent stack on their own hardware, that constraint is the whole story. No SSH, no systemctl status, no log file to grep at 2am when a task silently stalls. You get a UI and whatever telemetry OpenAI decides to surface. If that sounds limiting, it is, but it is also the trade most people will happily make.
A dot reaches the plugins you have connected across 4,000+ apps, and it can kick off work inside Codex and ChatGPT Work for research, analysis, documents and software. You talk to it through ChatGPT on web, desktop and mobile, plus Slack and Teams. Texting is listed as coming. There is no dedicated dot app.
The control layer is where I would spend my first hour. Custom Rules let you allow, block or require approval for specific actions. Auto-review inspects anything that could touch your accounts or share information outward. Activity View shows what the dot has been doing, including background runs it started while you were logged off. Background mode is read-only by design, so a dot browsing on its own cannot send a message or change content. Credentials get used through saved passwords without being handed to the model, and sensitive steps like a password change stay with you. Monitoring can pause or stop a dot if something looks wrong.
The timing of the safety messaging matters. One day before launch, OpenAI said its agents had posted user images online, affecting 53 ChatGPT users. OpenAI’s own safety post notes that dots still make mistakes and that consequential work needs review. I would not treat “read-only” as a guarantee that nothing surprising happens with your data.
What OpenAI dots mean in practice
What is an OpenAI dot?
A dot is a persistent agent inside ChatGPT that holds a standing goal instead of waiting for your next prompt. It gets its own cloud computer, its own browser, one identity and a name you choose. You talk to it on web, desktop, mobile, Slack or Teams, and it keeps working after you close the laptop.
How much do OpenAI dots cost?
The first dot is included with Pro and Business Premium at no extra charge. There is no standalone dot subscription to buy. OpenAI added a $500 tier the same day as the launch, and the ChatGPT pricing page is the only place worth checking for what your plan currently allows.
Do conversations with a dot count against my ChatGPT limits?
Chatting with your dot does not burn usage. Anything it kicks off inside Codex or ChatGPT Work does count, and that is the detail most people will miss. A dot quietly spawning long Codex runs spends your quota, not its own.
Can a dot act without asking me first?
Yes, within whatever Custom Rules you set. Auto-review and the safety monitor add a second gate on actions that touch accounts or share data. Set the default to require approval for anything outbound and you have effectively made it ask every time.
How do OpenAI dots compare to Meta Muse and Instinct?
Muse chases mass consumer reach and launched September 8, 2026 on Muse Spark inside a Secure VM, with a separate Sentinel agent approving outbound actions. Instinct lives inside text messages and has stayed invite-only since August 2026. Dots start with high-end subscribers and lean developer, which is a different bet on the same problem.
Specialist dots are the part aimed at companies rather than individuals. Each gets a fixed role, its own identity, its own credentials and access to company systems. OpenAI tested them internally across procurement, invoice processing, email marketing, customer support and commercial contracting, and rollout starts with enterprise pilots rather than open signup. The Microsoft Agent 365 work is the enterprise thread I would keep an eye on, because cross-vendor agent identity and permissions have been tangled every time anyone has tried it. That is not a reason to avoid the pilots. It is a reason to read the permission model before you connect anything with write access.
What to expect next
Texting is the surface OpenAI has already flagged as coming soon, and I think it changes the shape of the product more than Slack or Teams support did. Slack sits inside a workplace where you are at a keyboard and can read a long agent reply without much friction. A text message arrives when you are standing in a queue at the bank, and whatever the dot says has to be short enough to act on from a phone. I do not know how well a background research task compresses down to that. My expectation, and it is only an expectation, is that texting pushes dots toward the small jobs (status checks, quick approvals, “the invoice run finished”) while the long work stays on the desktop.
More eligible markets will follow the Pro and Business Premium rollout, and Enterprise, Edu and Healthcare workspaces stay behind an admin toggle until OpenAI widens the beta. Specialist dots are the piece I would watch most closely, because internal testing across procurement and invoice processing is a very different claim from a paying pilot where an accounting team has to trust the output.
That leads to the measurement problem nobody has solved. Suppose a dot spends four hours of background time reading supplier pages and hands you a two paragraph note. Was that worth a dedicated cloud computer? Nobody has given me a number I can apply to that question, and OpenAI’s own framing does not answer it. Anyone running a hosting business will recognise the shape of this, because it is the same argument we have about a VPS that sits at nine percent CPU all month. You keep it because it does real work at unpredictable hours, not because you can point at a utilisation graph and feel good. With a dot you cannot even see the utilisation graph, which is exactly why the Activity View matters. I do not know how this holds up once a single account runs five or ten dots, all doing background reading, all technically within quota until one of them opens a Codex task.
I would apply the same skepticism to the Microsoft Agent 365 partnership. Cross-vendor agent identity and permissions have been messy every time they have been attempted, and “specialist dots in Agent 365” is a sentence with a lot of unnamed work hiding inside it. It might land cleanly. I would not build a procurement workflow around it until the permission model is documented.
Your next step with OpenAI dots
If you are on Pro or Business Premium, the temptation is to hand your first dot a broad mandate and watch what it does. That is the wrong first move. Pick one job with a small blast radius and a clear failure mode. Triaging a shared inbox works well: the dot reads incoming mail, labels it, and drafts replies you approve before anything leaves. A weekly report built from a spreadsheet you already maintain works too, and it is easier to judge because you will notice immediately if the numbers are wrong.
Set Custom Rules before you write the goal, not after. Require approval for anything outbound, which covers sending, posting, submitting a form, or touching a connected account. Background mode keeps connected apps read-only, so a dot cannot send or change content while you are away, but that protection only applies to the background pass. Once you are in a conversation and ask it to act, the rules you configured are the only thing standing between a draft and a sent message.
Then let it run once and go read the aftermath. Open the dot’s computer from ChatGPT and pull up the Activity View, ideally the morning after its first background session. You want to see which plugins it reached for, how many steps it took to get nowhere useful, and where it stopped. That single log tells you more about your dot’s habits than the launch writeup (opens in new tab) does. I have found more bad automation behaviour reading a log at 8am than watching a task run live, because the live view only shows you the path it chose, not the ones it rejected.
Watch Codex separately. Conversations with a dot do not count against your ChatGPT usage, but anything the dot kicks off inside Codex or ChatGPT Work does, and that is where a quiet month can turn expensive.
On Enterprise, Edu or Healthcare, your first action is an email to whoever administers the workspace asking whether the dots beta is switched on, because nobody outside an admin console can turn it on for you. Agree the approval rules in writing before anyone connects a plugin with write access to a company system. Then create the dot, give it one sentence as a goal, and leave the rest alone until you have read its first background run.