ChatGPT dot is a continuous AI agent that keeps working on tasks between your conversations. The first dot is included at no extra charge with Pro and Business Premium. I had dot build a small to-do app. It could generate the app and add features, but dot did not finish checking the screen behavior, so I filled that gap in Codex. This article covers how dot differs from asking Codex directly, pricing, memory and permissions, and how to stop work.
What I learned from testing
I submitted a single-HTML app. A follow-up request also added "undo deletion."
The file would not open in dot's browser, so I tested the received file by operating it in Codex.
I did not use delegation to Codex. The official explanation introduces how delegation works.
What this article covers
- Codex direct, dot alone, and dot delegating: the differences
- Pricing, usage limits, and the first month
- Thinking about memory and permissions separately
- Initial setup and phone, Slack, and Teams
- Testing generation and fixes for a to-do app
- Things to watch for when delegating to Codex
- To stop, check the main dot, delegated work, and schedules
- Which to choose for your use case
- Frequently asked questions
- Related articles
Codex direct, dot alone, and dot delegating: the differences
According to the OpenAI dot overview, dot draws on a family of OpenAI frontier models, using different ones for different parts of its work, and has its own cloud computer and browser. GPT-6.1 Terra has now joined those models (see "What Is GPT-6.1 Terra? Pricing, Performance, and How to Choose Between It, Sol, Astra and Claude"). It covers research, documents, data analysis, and software creation. It keeps working even when you are not chatting, and it can coordinate multiple tasks. For app development, there are three routes:
You → Codex → Deliverable
Send the spec and fixes directly to a development chat. You can also ask for requirements work, tests, and review.
You → dot → Deliverable
A route where dot creates the work in its own workspace. The to-do app this time falls into this category.
You → dot → Codex
dot checks the results and gives follow-up instructions
A route where you hand off coordination of the work to dot. It requires the delegated environment and access permissions.
If you only need to fix one bug, telling an existing Codex chat the completion criteria is enough. dot is better suited when you want it to track the status of several problems, update investigation results and priorities, and let you know when a decision is needed. Rather than whether you can write code, the practical question is how much of the work of tracking multiple tasks and linking them to the next fix you want to hand off.
| Comparison point | Codex direct | Using dot |
|---|---|---|
| Daily instructions | Tell the development chat what to change while reviewing diffs and screens | Tell dot the goal and priorities, and open the delegated destination too if needed |
| Ongoing management | Continue the work in the dedicated chat | Coordinate multiple tasks and delegated destinations in the same conversation |
| Checking completion | Check the test output and the deliverable | Check the range dot verified and the range you need to supplement |
| After you turn off your PC | Cloud can keep running Local work needs the device | dot itself runs in the cloud Steps that use your local PC need the device |
Working while the PC is off is not unique to dot. Codex Cloud can also keep working while your PC is asleep.
Anthropic also began a trial release of Claude Duet, an app with a similar agent approach, in the United States on October 21, 2026. How it differs from dot is compared in What is Claude Duet?
Pricing, usage limits, and the first month
According to the OpenAI official help, the first dot is included at no extra charge with Pro or Business Premium, and there is also a quota for deep work (expanded in the launch month). This does not mean dot is available on the free plan.
You do not need a separate subscription to use your first dot with Pro. Conversations with dot do not count toward ChatGPT usage limits, and deep work has a quota included in your plan.
Work/Codex tasks started by dot also use your regular usage quota. Delegating alone does not add a charge.
Buying extra credits in Work/Codex is a separate payment. Using credits you already have and buying new credits are different things.
| Item | Official guidance |
|---|---|
| Individuals | Rolling out in stages to Pro 100, 200, and 500 Age and region requirements apply |
| Organizations | Business Premium and Enterprise Enterprise requires activation by an administrator |
| First dot | Included at no extra charge with Pro or Business Premium |
| Free, Go, Plus | No dot usage quota or trial count is listed |
| Conversations with dot | Not counted toward ChatGPT usage limits |
| Deep work | Quota included in your plan Specific numbers or token counts are not published |
| Delegating to Work/Codex | Counted toward the product's usage quota as usual |
Sources: dot access terms and the official help above (confirmed October 1, 2026, rechecked October 6). Pro is available except in the EEA, the UK, and Switzerland. Learn says the age requirement is over 18, while the release notes say 18 or older, so if you are near that boundary, check the guidance in your account.
The September 29 release notes say that for the next month, dot usage on eligible plans will not count against plan quotas. However, the dedicated guide says work started and managed in Work/Codex counts as usual, so the safest assumption is that delegated work uses your quota even in the first month. This does not mean one month from each person's creation date. A month later, on October 29, 2026 (UTC), OpenAI's Tibo wrote in a post on X that dot's expanded usage limits will continue for the time being and that dot usage is counted separately from regular usage limits. No end date was given. You can check usage in Work/Codex pricing and usage limits and in your account's usage status.
Thinking about memory and permissions separately
When I first talked with dot, it suggested a personal schedule that I had previously discussed with ChatGPT under the same account. As the official memory explanation describes, dot starts from related ChatGPT memory and also saves its own notes. When I corrected it with "Don't use earlier personal context. Research only public information," dot replied that it would correct its scope. However, narrowing the scope with an instruction does not erase information that has already been saved.
Context used for the first suggestion. It is separate from PC and Google connections.
Saves ongoing decisions and tasks. Changing ChatGPT settings does not necessarily delete them.
What it can read in apps and on your PC, and what it can operate.
Disconnecting does not erase information that has already been gathered. According to the official help, deleting dot's own memory requires deleting dot (Learn calls this Delete, and the help center calls it Reset). Save any deliverables before you make changes.
For proactive research, dot looks up information it is allowed to read and leaves what it finds in its notes. This research itself does not send messages, rewrite apps, or operate the browser or PC. The data settings explanation says that proactive research and private notes are not used directly for model training, but information that moves from there into conversations or tasks is subject to the data settings.
When you want to check before sending or publishing
The official Custom rules steps say to open "Settings → Personalization → Permissions → Custom rules" and use Add to specify actions and how they are handled. To require confirmation before sending or publishing, choose "Ask before taking action." For actions you want to perform yourself, choose "Hand off to you." As the official controls explanation says, rules are also instructions. Whether dot runs something automatically, waits for approval, or hands it off is decided from the instructions, permissions, and safety requirements. To keep information from being read, narrow it through connection permissions rather than through rules.
Initial setup and phone, Slack, and Teams
You can create a dot from the ChatGPT mobile app, the desktop app, or a PC browser (mobile web is not supported). Rollout is gradual, so the entry point may not appear right away even on an eligible plan. The official getting started guide says you can add app connections later and choose whether to connect your local PC. In my Windows app, I proceeded in this order:
- From "Your dot" to the introduction screen. "Create your dot" appeared. I read the explanations of ongoing work and data settings, then continued.
- Choosing a Google connection. There was an explanation of Mail, Calendar, Contacts, and Drive, along with an option to continue without connecting.
- Choosing where work runs. "Your dot's computer" and "This computer" for my local machine were shown. I continued without allowing my local PC.
- First conversation. dot's greeting, a notice about renaming it, and its first suggestion appeared.
Starting with public-information research or a small app, and adding connections only when you need them, is enough. I also received an HTML deliverable without connecting my local PC.
Skipping the Google connection does not disable plugins that are already enabled. As the official connections explanation says, dot can use the supported plugins enabled for your account within your existing permissions. To use an allowed local PC, the ChatGPT app must be open on that PC and online. The "Offline" label does not mean the permission has been revoked.
You can open the cloud PC from "Computers" in your profile. Choose Take over to operate it yourself, and Return control to hand it back to dot. The cloud browser has a separate login state from your local one, so sign in through a dedicated private form or a screen you have taken over yourself. Do not paste passwords into the conversation (official steps).
Using dot from your phone, Slack, and Teams
With the October 5–9 update, you can now create a dot from the phone app and set its name, appearance, and plugins. After creation, you can talk to the same dot from other devices.
Available in the mobile app, the desktop app, and the PC browser. You can also talk by voice using the phone button. Mobile web is not supported.
A contact route for exchanging messages with the same dot. Connect your contact in your profile. Read and write permissions for the app are separate.
The official explanation of contact methods says this is coming soon. Calls started from dot are also planned for after launch.
According to the official explanation of contact methods, it is the same dot whether you talk to it in ChatGPT, Slack, or Teams, but not every message is synced to each screen. From your phone, you can ask for a continuation of an existing local Codex task. However, execution stays local, so the ChatGPT app must remain open on the connected PC.
Adding dot to Slack alone does not start monitoring. Event monitoring requires specifying the target channel, the events, and the notification conditions, and it is only available when the connected service supports it. Here is an example request:
Example: Please check only the specified connected channel, on weekdays at 9:00 a.m. local time, for two weeks. Let me know in ChatGPT only about changes that affect deadlines or that need my decision. Limit replies to drafts, and do not send, publish, or purchase anything. Confirm the saved target, times, end date, and notification recipient.
After you register, you can check the status of the scheduled items in Scheduled.
Testing generation and fixes for a to-do app
On October 1, 2026, I (a Pro user) asked dot in the Windows Codex app to build a small to-do app. I specified that it should not delegate to a separate Codex, connect a local PC, or publish, push, or purchase anything. The request was along these lines:
Please build a small to-do app using only HTML, CSS, and JavaScript in your own cloud. It should support adding tasks, toggling completion, deleting, browser storage, and rejecting empty input. Report separately the operations you actually verified and the items you could not verify. Also explain how to receive the deliverable. Do not delegate to another development chat, connect PCs or external apps, or publish, push, or purchase anything. Also do not use earlier personal context.
dot submitted todo.html. HTML, CSS, and JavaScript were all in one file, and it implemented adding, completing, deleting, saving with localStorage, rejecting empty input, and a warning when saving fails.
dot's browser cannot open localhost
dot's cloud browser cannot open localhost, and HTML files are not allowed to be opened directly (October 1, 2026, dot in the Windows Codex app). dot stops at simulated tests, so for on-screen behavior, serve the received file from a temporary local HTTP server on your own machine and open it in Codex's built-in browser or similar to check. In this article too, I received the file and had a separate Codex chat operate it in the built-in browser.
I requested the spec → dot submitted the first HTML version
Simulated tests passed / could not be verified in an actual browser
I passed the file → operated it in Codex
I requested undo deletion → dot made a fix → re-checked in Codex
I added undo deletion
After checking the first version, I added a follow-up request: "Make it possible to undo deleting the item I just removed." In dot's revised version, an "Undo deletion" button appears after a deletion, and the item can be restored to its original position and completion state. When several items are deleted in a row, only the last one can be restored. Undo is also unavailable after a reload, and the screen explains this. Since dot again could not verify the revised version in an actual browser, I checked it in Codex.
Task deleted
You can undo the most recent one
Only the most recent one. Undo is not available after a reload.
What I confirmed in an actual browser
| Check item | Result of operations in Codex | Notes |
|---|---|---|
| Adding and blank input | Adding Japanese text succeeded Only blank input was rejected | Adding with Enter also worked |
| Completing and deleting | Toggling completion and deleting only the middle item worked correctly | Toggling completion with Space was also confirmed |
| Saving after a reload | The remaining items and their completion state were kept | localStorage when opened over HTTP |
| Undo | Restored to the original position and completion state | The restored result also persisted after a reload |
| Consecutive deletions and restoring after adding | Only the last deletion was restored The order with added items was correct | The undo button is hidden after restoring |
| Same name and last item | Restoring a completed item with the same name and restoring from an empty list were correct | Undo with Enter was also confirmed |
| Long text and screen width | The first version had no horizontal overflow at 390px or 1280px | — |
Input that looked like HTML tags was displayed as plain text. User input was inserted with textContent, and the code included no external scripts, images, or network communication.
Things to watch for when delegating to Codex
According to the official documentation, dot can create new cloud chats or Work/Codex chats on a connected PC, and it can also continue existing local Codex work. dot can review the results of delegated work and send follow-up instructions. You can also open the delegated work from "Activity" in your profile and give instructions directly.
However, not everything in your conversation with dot is handed over to a new delegated destination. You need to provide the goal, the completion criteria, prohibited actions, and any necessary background. Local files, browser login state, and uncommitted changes do not automatically move to the cloud environment.
To delegate development to the cloud, you first need to prepare a Codex Cloud environment with the target repository and development settings. In my test, the list of Codex Cloud environments visible from dot was empty. Also, the Codex Cloud limitations list "Computer and browser use" as currently unsupported. This means delegated work cannot be checked on screen the way I did here. If you have dot build an app, it is important to decide where you will verify completion first.
To stop, check the main dot, delegated work, and schedules
Ending a conversation or call does not stop the work you requested. The official stop steps have you operate three things separately:
- Pause the main dot. This stops the main dot's current work. It does not stop delegated work or future scheduled runs.
- Check delegated work in Activity. Open the development chat or task you want to stop and stop it individually.
- Check schedules in Scheduled. Disable or delete future scheduled runs.
Stopping does not undo completed changes, sends, or publishing. When you ask for ongoing work, deciding these four points at the start makes it clearer what needs to stop:
- Scope: this repository, this file, or only this research
- Completion criteria: what must work for it to be done, and which operation will verify it
- Permission scope: how far editing, delegation, sending, publishing, and purchasing may go
- Continuation conditions: whether it is one-time only, how long to keep tracking, and what to notify you about
Which to choose for your use case
For a small app, dot handled both the implementation and the follow-up fixes. However, it did not finish screen verification on its side. I filled that gap by handing over the file and checking in Codex. At this stage, I cannot say that letting dot handle the work removes the need to confirm completion yourself.
- An existing development chat already has the background
- You want to fix one repository while reviewing diffs and screens
- You want to decide the fix priorities yourself in detail
- You want to track research, documents, and schedules that are not about development, all together
- You want help organizing the status of multiple tasks and the next decisions
- You can prepare a delegated environment where behavior can be verified
Start with a small request, have dot report the deliverable, the operations it verified, and the items it could not verify, and then widen the scope of ongoing management you hand over.
Premise of this article: my test on October 1, 2026 covered only the route where dot itself builds the app. Delegation to Codex, phone, Slack and Teams, and the stop steps are based on official explanations (confirmed October 14, 2026).
Frequently asked questions
Is there a free trial for dot?
The official guidance does not list a dot trial count for Free, Go, or Plus. Your first dot is included at no extra charge with Pro or Business Premium. Rollout is gradual, so it may not appear right away even on an eligible plan.
Is dot cloud-only? Can't I build an app without connecting a PC?
The dot itself runs in the cloud, and an allowed local PC can also be used for work. The single HTML file this time was built in dot's own cloud, and no PC connection was needed. If you want dot to handle local files or an existing local Codex, a connection is required, and that PC must be online with the ChatGPT app open. Cloud work continues even when your PC is off (official computer explanation).
What is the Start call button for?
It is the entry point for talking to dot by voice. According to the official overview, you can start a call from the phone button in a conversation, and you can also send text during the call. Ending the call does not stop the work you delegated.
Will asking dot give better quality and lower cost than Codex?
I did not make a comparison under the same conditions. What I found is that dot could handle generating and fixing a small app, but screen verification had to be arranged separately.
Related articles
- ChatGPT Pro 100, 200, and 500 comparison: how to choose the Pro plan that includes dot
- GPT-6.1 Sol vs. 6 Sol comparison: how to choose a development model for Codex
- How to keep completion criteria while working with Codex /goal: managing a goal in a single Codex chat
- Differences when closing your PC with Remote, Cloud, and Work: where work runs and PC requirements
- How to run scheduled tasks in ChatGPT: scheduling one-time notifications or daily information summaries
- How to hide the ChatGPT and Codex pets, and when to use Mini: removing the character floating on screen (tasks keep running)