ChatGPT dot is an ongoing AI agent that can keep working between conversations. How does it differ from asking Codex directly? We tried the setup flow, had dot create and modify a ToDo app, and examined pricing and the work you can entrust to it.
What this test established
dot delivered a single-file HTML app and added an undo-delete feature when requested.
dot did not complete browser verification. The received files were operated and checked in Codex.
dot built the app itself. We did not measure delegation to Codex or compare quality and usage.
What this guide covers
- Codex directly, dot working alone, and delegation through dot
- Pricing, free access, and the first-month offer
- Memory and permissions are separate
- The setup flow and connection choices
- Creating and modifying a ToDo app
- What we verified in a real browser
- Requirements for delegation and independent checks
- Stopping dot, delegated work, and scheduled runs
- Which approach fits your needs?
- Frequently asked questions
Codex directly, dot working alone, and delegation through dot
OpenAI’s dot overview describes an agent powered by GPT-6 Astra, with its own cloud computer and browser. Its scope includes research, documents, data analysis, and software creation, as well as working between conversations and coordinating multiple responsibilities. For app development, it helps to distinguish these three routes.
You → Codex → Deliverable
Give specifications and revision requests directly to the development chat. You can also request requirements planning, testing, and review.
You → dot → Deliverable
The work happens in dot’s workspace. This is the route used in our ToDo test.
You → dot → Codex
dot reviews the result and gives follow-up instructions
dot coordinates the work. This requires an available development environment and access permissions.
The distinction is ongoing coordination more than the initial coding request
For a single bug, you can give acceptance criteria to an existing Codex chat and review its test results. Tracking several issues, updating research findings and development priorities, and notifying you when a decision is needed are examples where dot’s ongoing coordination may be worth considering. We did not test that latter workflow.
Asking Codex directly can also cover acceptance criteria and rounds of revisions, beyond implementation itself. To assess dot’s value, it is more useful to ask how much responsibility you want it to take for tracking multiple jobs, checking results, and arranging the next fixes than simply whether it can write code. Our test provides no evidence that using dot necessarily produces better code or consumes less of your allowance.
| Point of comparison | Codex directly | Using dot |
|---|---|---|
| Day-to-day instructions | Review changes and screens in the development chat, then give instructions | Give dot goals and priorities; open delegated chats when needed |
| Ongoing coordination | Continue the work in the responsible chat | The official guide describes coordinating multiple responsibilities and delegated tasks in one conversation |
| Checking completion | Review test output and deliverables | Establish what dot checked and what still needs human verification |
| After your PC is turned off | Cloud work can continue Local work needs the computer | dot itself runs in the cloud Steps using your PC need that computer |
Working while your PC is off is not unique to dot. The official Codex Cloud documentation also describes work continuing while your PC sleeps. Either approach may need your attention for authentication, approval, errors, or other interruptions.
Pricing, free access, and the first-month offer
The official OpenAI help guide says your first dot is included with Pro or Business Premium at no additional charge. There is also an allowance for deeper work, expanded during the launch month. This means inclusion in an eligible paid plan, rather than availability on a free plan.
| Item | What the official sources establish |
|---|---|
| Personal plans | Rolling out to Pro 100, 200, and 500 Age and regional conditions apply |
| Organization plans | Business Premium and Enterprise Enterprise requires administrator enablement |
| Your first dot | Included with Pro or Business Premium at no additional charge |
| Trials on Free, Go, or Plus | The guidance we checked gives no dot allowance or trial count for these plans |
| Conversations with dot | The dedicated guide says these do not count toward ChatGPT usage limits |
| Deeper work | An allowance is included in the plan Exact task or token counts could not be established |
| Delegation to Work or Codex | The dedicated guide says it counts toward the product’s allowance as usual See below for uncertainty about the first-month offer |
Sources: dot access conditions and the help guide above, checked October 1, 2026. Pro availability excludes the EEA, UK, and Switzerland. Learn says “over 18,” while the release notes say “18 and older”; check the account’s access guidance for eligibility at that boundary.
The first month does not establish that everything is free—or that delegation always counts
The September 29 release notes say dots usage will not count toward eligible plan allowances for the next month. The dedicated guide, however, says work started or managed in Work or Codex counts as usual. These statements could be compatible if dot itself and delegated work have different scopes, but the original wording does not settle how delegation or environment preparation is treated.
We did not measure changes in usage or billed amounts. Do not treat the first-month offer as a guarantee that all work is free and unlimited. It also does not waive payments to external paid services. The announcement says conditions after the first month will be communicated later; do not reinterpret it as a full month starting on each user’s creation date.
More delegation steps could mean more processing for additional reviews and retests. Better requirements planning might reduce rework. These are operational possibilities, rather than a measured fixed surcharge or savings rate for using dot. Check Work and Codex pricing and allowances and your account’s usage, distinguishing consumption of existing credits from new purchases or automatic top-up charges. For choosing Pro itself, see our Pro 100, 200, and 500 comparison.
Memory and permissions are separate
In the first conversation, dot suggested a private plan previously discussed in ChatGPT under the same login. This happened before the user entered that plan in the dot conversation, making it an observed instance of earlier context appearing in the first suggestion. We are not publishing the specific plan or anyone’s name.
The official memory explanation says dot starts with relevant ChatGPT memories and also saves its own notes. Our observation is consistent with that description, but we cannot identify which saved memories or past conversations were used. It is not evidence of information leaking from another account or of the entire chat history being copied.
Context relevant to initial suggestions, separate from PC or Google connections.
Stored decisions and ongoing work. Changing ChatGPT settings does not necessarily delete them.
What dot can read or operate in apps and on a PC.
The user corrected dot: the plan was not part of their work, and dot should research only public OpenAI information without using previous private context. dot acknowledged the correction and said it would change the scope. We do not treat that reply as proof that stored information was deleted or all related tasks were stopped.
Disconnecting an app is also different from deleting information already obtained. The official help guide says deleting dot is necessary to remove its own saved memories. Learn uses the name Delete, while the help guide uses Reset, so read the confirmation shown in the actual interface before deciding. The user should save any deliverables and check what will be removed before taking that action.
Instructions and custom rules such as “ask before sending” or “do not publish” are useful, but they are not settings that revoke access, nor a guarantee against mistakes. The official controls guide says instructions, existing permissions, and safety requirements determine whether an action runs automatically, waits for approval, or is handed back to you. It does not say every action requires confirmation every time.
If you want confirmation before sending or publishing
The official Custom rules instructions describe opening “Settings → Personalization → Permissions → Custom rules” after setup, then using Add to specify an action and how it should be handled. Choose “Ask before taking action” for actions such as sending or publishing that you want to approve first, or “Hand off to you” for actions you always want to perform yourself. Availability and display depend on the account and administrator settings.
Define the scope, acceptance criteria, and actions that are prohibited for this request.
Check ongoing confirmation rules and permissions to connect to apps and your PC.
This is an explanation of the official procedure; we did not change these settings and test their behavior in this experiment. Custom rules are also treated as instructions. They are not equivalent to revoking connection permissions or completely preventing unintended actions.
The setup flow and connection choices
You create dot from ChatGPT in the desktop app or a desktop browser. The official getting-started guide says you can skip app connections and add them later, and choose whether to connect your PC. In our Windows app, the screens appeared in this order.
- Open “Your dot” to reach the introduction. The screen showed “Create your dot.” The user read the explanations of ongoing work and data settings, then continued.
- Choose whether to connect Google. Mail, calendar, contacts, and drive were listed, with an option to continue without connecting.
- Choose where dot can work. The screen showed “Your dot’s computer” and an option for the local computer. For this test, the user did not authorize access to the local PC.
- Enter the first conversation. dot displayed a greeting, an invitation to change its name, and its first suggestion.
Begin with public-information research or a small app, then consider connections as they become necessary. We received an HTML deliverable without connecting the local PC.
Skipping Google connections does not disable plugins that are already enabled. The official connection guide says dot can use supported plugins enabled on the account within their existing permissions. Before starting, check which account is logged in, which apps are connected, and their read and send permissions.
In the profile’s “Computers” section, distinguish the cloud computer from your local PC. The presence of a button to authorize access does not mean access has been authorized. To use an authorized PC, it must be online with the ChatGPT app open. An “Offline” status also does not mean access permissions have been revoked.
A small first request without connections
For an initial test, keeping the deliverable and verification method narrow makes the result easier to assess. The following example adapts our request for readers; we did not run this exact wording as a comparative experiment.
Create a small ToDo app in your own cloud using only HTML, CSS, and JavaScript. Support adding tasks, toggling completion, deletion, browser storage, and rejection of empty input. Report the operations you actually verified separately from anything unverified, and explain how I can receive the deliverable. Do not delegate to another development chat, connect to my PC or external apps, publish, push, or make purchases. Do not use previous private context.
In the completion report, check not just what was implemented, but whether actual operations were performed or only simulated tests, and which file was checked. An instruction not to use private context is not a setting that deletes stored information.
Creating and modifying a ToDo app
Test conditions: On October 1, 2026, a Pro subscriber made the requests to dot in the Windows Codex app. OpenAI describes dot as powered by GPT-6 Astra, but we did not obtain the execution model for this work or the client’s detailed version number. The request excluded delegation to another Codex task, local PC connections, external publication, push, and purchases. The target was a small ToDo app using only HTML, CSS, and JavaScript.
The request specified both features and verification
Key points of the request
- Build in dot’s own cloud without delegating to another Codex task
- Support adding tasks, toggling completion, deletion, and browser storage
- Actually operate the app in a browser to verify it
- Explain how to receive the deliverable and identify anything not verified
- Do not connect to the PC or external apps, publish, push, or make purchases
dot initially said it would create the app and check it in a browser, then delivered todo.html. HTML, CSS, and JavaScript were contained in one file. Inspection of the source found task addition, completion toggling, deletion, localStorage persistence, rejection of empty input, and a warning for storage failures.
dot delivered the app but could not complete browser verification
dot reported that its cloud browser was blocked from accessing localhost and was not permitted to open HTML directly. It said its simulated DOM and storage tests passed, but explicitly stated that operation, display, and persistence in a real browser were unverified. We did not obtain raw error logs or establish the cause of the access restrictions.
Simulated tests can help check logic, but they do not prove that storage, rendering, or keyboard operation works in an actual browser. At this point, the requested real-browser verification had not been completed. This also does not establish that all dot environments have the same restrictions.
The user requested the specifications → dot delivered the initial HTML file
dot reported passing simulated tests / real-browser behavior remained unverified
The user passed on the file → Codex operated it in a real browser
The user requested undo → dot modified the app → Codex tested it again
Adding undo for a deletion
After checking the initial version, the user requested the ability to undo the most recent deletion. This was an improvement outside the initial requirements, rather than a bug fix. In the revised file delivered by dot, an undo-delete button appeared after deletion, with logic to restore the item’s original position and completion state.
Task deleted
You can undo the most recent deletion
Only the most recent item. Undo is unavailable after reloading.
After consecutive deletions, only the last deleted item can be restored, and reloading clears the undo history. The interface explains this. dot reported passing simulated tests for position, completion state, consecutive deletions, and other cases, but said the revised version had also not been checked in a real browser. The user received the revised file and passed it to Codex for another check.
What we verified in a real browser
The received originals were served unchanged through a temporary local HTTP server, and the AI in a separate Codex chat operated the built-in browser. We recorded the SHA-256 hashes of the original and revised versions to identify exactly which files were tested. This was actual operation of the deliverables in another environment, rather than merely reading dot’s own report. It was not an assessment based on long-term everyday human use.
| Check | Actual operation results in Codex | Conditions and limitations |
|---|---|---|
| Addition and whitespace input | Japanese-language tasks were added successfully Whitespace-only input was rejected | Addition also worked with Enter |
| Completion and deletion | Completion toggling worked, as did deleting only the middle item | Completion toggling with Space was also checked |
| Persistence after reload | Remaining items and their completion states were retained | localStorage when opened over HTTP |
| Undo | The original position and completion state were restored | The restored result also persisted after reload |
| Consecutive deletions and restoring after addition | Only the last deletion was restored The order relative to newly added items was correct | The undo button disappeared after restoration |
| Identical names and the last item | A completed item sharing another item’s name and restoration from an empty list both worked | Undo with Enter was also checked |
| Long text and screen widths | The initial version had no horizontal overflow at 390px or 1280px | The revised version was not checked across all widths or zoom levels |
Source: AI Arte’s operation records from October 1, 2026. The claim that dot’s simulated tests passed came from the report provided by the user; we did not independently inspect the raw test logs.
Input resembling HTML tags was also displayed as text. The code inserted user input using textContent and contained no external scripts, images, or network operations. These checks alone do not guarantee the security of the entire app.
We did not check opening the HTML by double-clicking, changing the file’s name or location, private browsing, insufficient storage, or simultaneous editing in multiple tabs. Success over HTTP does not establish success for every way of opening the file.
A full task-management app would also need consideration of recovery from storage failures, synchronization between tabs, data export, and accidental deletion. Here we checked basic and added features in a single-screen prototype. We did not score its readiness as a product or dot’s ability to handle large-scale development.
Requirements for delegation and independent checks
The official documentation says dot can create new cloud chats or Work and Codex chats on a connected PC, and continue existing local Codex work. It can check results from work it started and send follow-up instructions. Users can also open delegated chats from “Activity” in the profile and give instructions directly.
A new delegated chat does not automatically receive the entire conversation with dot. It needs the goal, acceptance criteria, prohibited actions, and relevant background. Do not assume local files, browser login state, or uncommitted changes transfer automatically to a new cloud environment. Also distinguish a reply saying a model was specified from a record of the model that actually executed the work.
Preparing Codex Cloud alone does not guarantee browser inspection
We received a report that the Codex Cloud environment list available to dot was empty. Delegating cloud development requires preparing a Codex Cloud environment with the target repository and development configuration first. We did not create an environment or perform delegation in this test.
However, the limitations on that same official page explicitly list “Computer and browser use” as currently unsupported. We therefore cannot say that preparing an environment enables inspection with the built-in browser. We also did not test whether command-driven browser testing with tools such as Playwright works in this environment.
Read code, logs, and images to identify gaps. This is different from rerunning actual operations.
Run the same version separately, then enter input, click, and reload. A reachable access route is required.
dot replied that, with localhost and file URLs unavailable, it had not established a way to operate the app itself without publication or a PC connection. A delegated environment’s localhost is not automatically reachable from another cloud environment. We did not proceed with further preparation; instead, the user passed on the deliverables for operational checks in Codex.
This does not show that dot lacks delegation features. It means we had not secured a delegated environment or an independent inspection route under these test conditions. Distinguishing successful standalone app generation from successful coordination after delegation is essential when assessing practical usefulness.
Stopping dot, delegated work, and scheduled runs
Ending a conversation or voice call does not necessarily stop work already assigned. The official stopping instructions treat the following as separate actions.
- Pause dot itself. This stops its current work, rather than all delegated tasks and future recurring runs at once.
- Check delegated work in Activity. Open the development chats or jobs you want to stop and stop them individually.
- Check scheduled runs in Scheduled. Disable or delete future recurring runs. Stopping a running job and canceling future schedules are separate actions.
Stopping does not undo changes, messages, or publications already completed. We did not test this entire stopping flow, so the steps above are presented as the official procedure. When assigning ongoing work, define when it should end and where results should be reported, so the work to stop is easier to identify.
- Scope: only this repository, file, or research topic
- Acceptance criteria: what counts as finished and which operations will verify it
- Permissions: how far editing, delegation, sending, publication, and purchasing are authorized
- Ongoing work: whether this is a one-off request, how long to track it, and what to notify you about
Check both the written conditions and the actual state of connection settings, Activity, and Scheduled.
Which approach fits your needs?
For this small app, dot implemented the initial version and the requested modification. Browser verification was not completed on dot’s side: the user transferred the files and a separate Codex chat filled that gap. We cannot currently say that entrusting work to dot eliminates the effort of checking completion.
- An existing development chat already has the relevant background
- You want to modify one repository while reviewing changes and screens
- You want to set detailed revision priorities yourself
- You want to track research, documents, and plans alongside development
- You want help organizing several jobs and upcoming decisions
- You can provide an accessible delegated environment and verification route
This is a suggested division of roles based on the official descriptions and our observations, not a controlled experiment in which both approaches developed the same app. Start with a small request and ask for the deliverable, the operations actually checked, and anything unverified. Then decide whether to expand the scope of ongoing coordination. Choosing a Codex development model is a separate question from dot’s role; see our GPT-6.1 Sol versus 6 Sol comparison.
Frequently asked questions
Does dot have a free trial allowance?
The official guidance we checked does not specify a dot trial count for Free, Go, or Plus. Inclusion of the first dot in Pro or Business Premium at no additional charge is different from availability on a free plan. Access is rolling out gradually, so even eligible subscribers may not see it immediately.
Must I connect my PC to create an app?
The single-file HTML app in this test was created and modified in dot’s own cloud. A PC connection was not required. Handling files on your PC or working with an existing local Codex chat requires a separate connection.
What does the Start call button do?
It starts a voice conversation with dot. The official overview describes starting a call with the phone button in the conversation and sending text during the call. Ending the call does not stop previously assigned work.
Will dot produce better results than Codex at a lower cost?
We did not perform that comparison. Creation and modification worked, but independent browser inspection was performed in a separate Codex chat. Without recording time, human intervention, additional reviews, model settings, and usage under the same conditions, we cannot judge relative quality or cost.
