What are ChatGPT dots?

Dots are always-on AI agents in ChatGPT, announced by OpenAI on September 29, 2026. Each dot runs on GPT-6 Astra, has its own cloud computer and browser, works across the apps you connect, and can keep working on scheduled or ongoing tasks between your conversations. You can message it in ChatGPT, Slack or Microsoft Teams.

Who can use OpenAI dots?

As of September 30, 2026, dots are rolling out to ChatGPT Pro users outside the European Economic Area, Switzerland and the UK, and to Business Premium users in all supported ChatGPT regions. Enterprise, Edu and Healthcare workspaces can try a beta once an admin turns it on. Your first dot is included in the plan.

Create your dot in the ChatGPT desktop app or on desktop web, connect the apps it needs, then give it a goal and the details it needs. You can ask it to run recurring checks, set custom rules for which actions it may take alone, and follow its work in its profile under In progress, Scheduled and Completed.

Can a dot run A/B tests?

A dot can watch your analytics, draft a test brief and, through Codex, prepare a variant for review. Whether it can launch a test depends on the tools you connect. We recommend a custom rule that hands launching, stopping and promoting tests back to you, and judging every test on revenue per eligible visitor.

What can AI agents do for marketing?

Agents are useful for the repeated work between decisions: checking funnels and campaigns on a schedule, summarizing what changed, drafting briefs and variants, and preparing reports. They are weaker at deciding what is worth testing without evidence, and they cannot make an underpowered test conclusive.

OpenAI Dots: What They Change for A/B Testing and Analytics

OpenAI dots explained for growth teams and agencies: what the always-on agents can do, how to set approval rules for tests and spend, and how to judge whether a dot actually made you money.

Published Sep 30, 2026Updated Sep 30, 2026

OpenAI Dots: What They Change for A/B Testing and Analytics

The short answer: As of September 30, 2026, OpenAI dots are always-on GPT-6 Astra agents inside ChatGPT. Each has its own cloud computer, works across your connected apps, runs scheduled tasks, and follows approval rules you set. For teams running tests, a dot changes who watches the funnel between meetings and who drafts the next brief. It does not change what makes a test valid. Our call: give a dot the monitoring and drafting work now, and set rules so that launching a test, picking a winner and moving ad spend always come back to you.

TL;DR

Dots are rolling out to ChatGPT Pro (outside the EEA, Switzerland and the UK) and Business Premium, with an Enterprise beta.

The useful change for growth teams: a recurring, proactive check on your funnel that ends in a draft brief, not just a chat answer.

The risk: a dot keeps what it learned from a connected app even after you disconnect it. Agencies should think about client data before connecting anything.

A dot does not add conversions. If a test can't reach enough of them, a faster test cycle won't fix it.

What OpenAI shipped

OpenAI announced dots at DevDay on September 29, 2026. In OpenAI's words, they are "always-on agents built to handle everything." The facts that matter for a marketing team:

Model and computer. Each dot runs on GPT-6 Astra, OpenAI's most capable model, and has its own cloud computer and browser. You can open that computer and watch its work. Access to your own laptop is optional and off by default.
Apps. Dots connect to apps through ChatGPT's plugin ecosystem, which OpenAI says covers more than 4,000 apps. You manage access through your existing ChatGPT app controls.
Proactive work. When you aren't talking to it, a dot does what OpenAI calls proactive research: it reviews connected apps with tools restricted to read-only, forms memories, and suggests ways to help.
Scheduled tasks. You can ask a dot to run a recurring check and review those tasks under Scheduled in its profile.
Codex. A dot can create tasks in Codex cloud environments, and, with your computer connected, Codex tasks on your machine using your local skills.
Rules. Custom rules let you set, per action, one of four behaviors: take action without asking, take action if pre-approved, ask before taking action, or hand off to you. OpenAI also runs an automatic review of actions that could affect your accounts.
Where you talk to it. ChatGPT on desktop, web and mobile, plus Slack and Microsoft Teams. Texting is a limited US beta for Pro users.
Cost. Your first dot is included in Pro or Business Premium. Conversations with it don't count toward ChatGPT usage limits, but Codex and ChatGPT Work tasks it starts do.

What it changes for A/B testing

Area	Before	With a dot	What to check
Watching the funnel	Someone opens the dashboard when they remember to	A scheduled check runs whether or not anyone asks	That it compares the same date ranges and event definitions every time
Finding test ideas	Ideas come from meetings and gut feel	The dot proposes ideas from what it saw in connected data	That each idea cites the data it came from
Building variants	A developer picks up a ticket	The dot can start a Codex task and return a pull request for review	That the variant changes only what the brief says
Tracking and events	Checked by hand before launch	The dot can check events in its browser, if asked	Exposure and conversion fire once each, assignment is stable
Deciding winners	Analyst reads the result	The dot can summarize the result	Who is allowed to call it, and on which metric
Ad spend	Media buyer adjusts budgets	Depends on the connected ad tools and your rules	Hand budget changes off to a person

The pattern: dots make the repeated parts of a testing program cheaper. The judgment parts stay where they were.

Set up a dot for a testing program

Give it one goal that names the money metric. For example:

Your goal: help us find and run A/B tests that raise revenue per
eligible visitor on [site], for paid traffic from [channels].

Every Monday at 9am, compare last week with the four weeks before:
traffic, conversion and revenue by landing page, device and campaign.
Use the same date ranges, timezone and event definitions each time.

Flag a page only when the change is large enough to matter and the
numbers behind it are named. For one flagged page, draft a test brief:
audience, control, the single change, primary metric, guardrails,
and the evidence that would prove the idea wrong.

Never launch, stop or promote a test, change a live page, or change
an ad budget. Bring those to me.

Then set custom rules to match. A starting point:

Action	Rule
Read analytics, ad reports and docs	Take action without asking
Draft a test brief or report	Take action without asking
Start a Codex task to build a variant in a preview	Ask before taking action
Post a summary to the team's Slack channel	Take action if pre-approved
Launch, stop or promote a test	Hand off to you
Change a live page, ad budget or bid	Hand off to you
Send anything to a customer or client	Hand off to you

OpenAI's own help page notes that a dot "can make mistakes, including when following your rules." Treat the rules as a safety net, and keep the high-stakes actions in tools that require a person anyway.

What to check before you rely on it

What it keeps from connected apps. OpenAI's help center says disconnecting an app does not delete what your dot already learned from it; you have to reset the dot. If you are an agency, connecting one client's analytics to the same dot you use for another client mixes their context. Use a separate workspace per client, or don't connect client data until you've read the privacy FAQ.
Which plan and region you're on. Pro access excludes the EEA, Switzerland and the UK. Business Premium covers all supported regions. A team split across regions may not all get it.
Whether its numbers match your dashboard. Before trusting the Monday check, run it on a week you already analyzed and compare every figure.
Whether a variant it builds still tests your hypothesis. A pull request can pass review and still change pricing copy, consent behavior or the analytics ID. Confirm stable assignment, exposure when the change is actually seen, and conversions that survive a handoff to checkout.
Usage. Chatting with the dot is free of usage limits; the Codex and Work tasks it starts are not. A dot that builds a variant every week uses real quota.

How to measure whether it helped

A dot is worth keeping if it leads to more valid tests that move revenue, not if it produces more ideas. Track three things for a month before and a month after:

Tests launched that had a written primary metric and a feasibility check.
Time from "we noticed a drop" to a reviewed brief.
Test results scored on revenue per eligible visitor, not signups or clicks.

Check feasibility before any test, whoever proposed it. The conversions a test needs are driven by the smallest lift you want to detect: roughly 15.7 divided by the square of that lift, per variant. To detect a 10% relative lift you need around 1,570 conversions in each arm. Use the sample size calculator and the minimum detectable effect guide. If your site can't reach that in a few weeks, a dot that drafts tests faster won't help; it will help more with attribution and funnel monitoring.

Using dots with Humblytics

Humblytics does not have a ChatGPT plugin today, so you won't find it in your dot's plugin list.

The supported route is Codex. Humblytics' MCP server installs in Codex and gives an agent read access to traffic, funnels, pages, clicks, forms, split tests and Meta ad data, plus a narrow write surface: create, update or stop a split test and save a funnel. Every write is a proposal the user approves, and promoting a winner is always the user's decision. That lines up with the custom rules above.

Because a dot can start Codex tasks, a dot can hand the analysis step to a Codex session that has Humblytics connected. We have not yet confirmed that a Codex task started by a dot loads the MCP servers from your Codex setup. Test it first with a read-only request, such as asking the task to list your Humblytics properties, before you build a weekly check on it.

Revenue answers need Stripe or Foxy connected in Humblytics; ad attribution needs a connected Meta or Google Ads account as well.

Set up the Humblytics MCP in Codex or Claude Code, then point your dot at one funnel and one question.

For which model to use for each part of the work, see the best AI model for A/B testing.

Sources and freshness

OpenAI, Introducing dots, published September 29, 2026; retrieved September 30, 2026. Supports the description, GPT-6 Astra, cloud computer, 4,000+ apps, proactive research, custom rules, auto-review, Slack and Teams, plan availability and usage limits.
OpenAI Help Center, Getting started with your dot, retrieved September 30, 2026. Supports regional availability, setup steps, Codex tasks, scheduled tasks, the four rule behaviors, the note that disconnecting an app does not delete learned information, and the texting beta.
OpenAI Help Center, Dots privacy, security, and safety FAQs, retrieved September 30, 2026. Referenced for data handling.
Humblytics, MCP and skills, retrieved September 30, 2026. Supports the read and write scope, approval rules, and supported agent clients.
Conversion formula from our sample size calculator, two-sided test at 95% confidence and 80% power.

The setup, rules and prompts are proposed workflows, not measured customer results. Dots are rolling out gradually and OpenAI is changing them often; check OpenAI's pages before relying on a specific feature.