OpenAI introduces dots, an AI assistant for everyday work

Research a topic, compare options, and coordinate schedules. Update the slides when the numbers for a meeting change. Each task may be small, but remembering what comes next and keeping the relevant information together takes effort.

Announced by OpenAI on September 29, dots is a personal assistant designed to take on this kind of ongoing work. Powered by GPT-6 Astra, it has its own cloud computer and browser. OpenAI presents it as an agent that works between conversations and returns with results or matters that need a human decision. Official announcement

What would be useful to delegate? This article explains dots through the everyday and workplace scenarios OpenAI describes. Examples such as comparing venues and coordinating schedules are official illustrative scenarios. Later, we separately describe the author's use of Gmail and calendar connections, then consider the convenience of delegating assistant duties alongside the security and privacy conditions of granting access to information.

Key points

• The intended use goes beyond researching options: dots can update comparisons and documents when conditions change and prepare what is needed for the next decision.
• You can keep talking to dots while it works on several tasks. You set the goal and scope, and take part in decisions such as sharing or purchasing.
• In the cloud environment examined, we found executable files for 12 apps covering 3D, CAD, images, video, maps, and other work. They suggest broader creative and analytical uses, but their presence does not demonstrate successful production of finished work.
• The first dot is included with Pro and Business Premium at no additional charge. Work and Codex tasks started by dots still count toward their usual usage limits.
• When delegating email and calendar work, distinguish reading from sending or editing, and disconnecting an app from deleting existing data. Reporting by the AI itself is also separate from safety review or legally grounded disclosure by OpenAI as the service operator.

Sources for the first draft were checked on September 30, 2026. Permissions, privacy, reporting, availability, and pricing were rechecked, and the cloud environment examined, on October 1, 2026, Japan time. We distinguish official illustrative scenarios, the author's own use of Gmail and calendar connections with a request for assistant duties, and information read from the cloud environment examined. This article does not measure successful creative outputs or savings in time or money.

Contents

1. Keep talking while dots works on several tasks
2. From comparing options to scheduling and updating documents
3. Following up when conditions change
4. What the cloud's creative and analytical tools could make possible
5. You decide what dots can access and do
6. Delegating Gmail and calendar work also means deciding how data is handled
7. Would an AI report an ethically troubling email?
8. Eligible plans still face a gradual rollout and task usage limits

Keep talking while dots works on several tasks

OpenAI's product page describes dots as an agent that learns how you work and moves projects forward using Codex and connected tools. You provide a goal, and can add constraints or change priorities along the way.

When given several tasks, dots can delegate to other agents, review their results, and issue further instructions. You can continue talking to dots while the work is underway. Task guide

Continuity comes from conversation context, relevant ChatGPT memories, and notes dots keeps about preferences and decisions. This does not mean every conversation is recorded in full and can always be retrieved. The official guide says context is selected for each task; entire conversations are not automatically passed to separate tasks. Memory and context guide

The aim is to reduce the need to explain everything from scratch and connect new information to ongoing work. OpenAI explicitly identifies GPT-6 Astra as the model behind dots, but its unpublished internal architecture cannot be used to explain these behaviors.

From comparing options to scheduling and updating documents

Grouping the official examples by the outcome a user wants helps clarify dots' role. The human involvement column below summarizes the constraints and approvals in those examples. It does not guarantee that the same operations are available in every service.

Official scenario	What dots works on	Where the user is involved
Finding a venue for a gathering	Compare capacity, travel, and pricing	Specify a budget and other requirements
Resolving a scheduling conflict	Find another time when participants can attend	Approve moving the event
Keeping meeting materials current	Check sales or usage figures and update the materials	Review the content and approve sharing
Choosing a meal order	Present options with total cost and delivery time	Choose an option and approve the order

Sources: ChatGPT Learn's venue comparison example and the scheduling, document update, and meal ordering examples on the dots product page. Available services vary by region and connection. We have not verified support for individual services in Japan.

The useful point in the venue example is that dots assembles information for a comparison on common terms, rather than merely naming candidates. Seeing capacity, travel, and price together gives a person a basis for choosing. The official introduction also shows a user adding a budget ceiling while a comparison is in progress. Introduction guide

At work, maintaining existing documents is an easy use case to picture. The announcement describes updating a proposal and validation plan as a customer's requirements and seat count change during a sales process, then bringing unresolved questions back to the person responsible. Each change prompts another pass through the relevant documents and remaining work. Sales support example in the announcement

These examples share an emphasis on preparation before the user decides: read and compare materials, investigate the consequences of a change, and present an update. Any subsequent sharing or ordering depends on the scope of the request and the required approvals.

Following up when conditions change

A comparison table or plan becomes outdated when headcounts, prices, or deadlines change. The official description of dots emphasizes continuing the work in response to those changes.

For example, when preparing a gathering, dots is shown updating the headcount as people reply and checking with the user before requesting a revised quote. The intended workflow connects the initial research result to the next follow-up. Ongoing work example in the introduction

A clearer brief specifies not just what to research, but which changes should trigger an update and which decisions should prompt a notification. For recurring tasks, the official guide likewise recommends specifying the subject, execution time, notification conditions, and delivery destination, then checking the saved schedule. Recurring task guide

Following up on ongoing work is different from repeating a task at a fixed time. The latter requires a saved schedule. Supported external services can also trigger tasks when particular events occur, but connecting an app does not automatically create recurring or event-triggered tasks.

OpenAI also describes proactive research as a way for dots to read authorized information and look for useful next steps. The research tools themselves do not directly contact other people or modify connected services. Any action taken afterward remains subject to the permissions and approvals for that action. Official safety explanation

What the cloud's creative and analytical tools could make possible

The computer that dots has at its disposal offers more than a browser for research. OpenAI describes an environment built on Linux and Chrome for creating files and running tools and code. Work can continue while your own device is switched off, and you can take over the screen controls when needed. Cloud environment explanation and computer guide

In the dots cloud environment used for this investigation, we confirmed that executable files for the following 12 apps were present. This finding comes from reading the actual environment, independently of whether public documentation lists the software. Blender reported version 4.3.2, and Godot reported 4.6.3 stable. We did not, however, verify successful use of each app's interface, creation of finished work, rendering, or export.

Use	Apps whose executable files were found
3D modeling and animation	Blender
CAD and solid modeling	FreeCAD、OpenSCAD
Electronic circuit and PCB design	KiCad
Editing photos, illustrations, and images	GIMP、Inkscape、ImageMagick
Video editing	Kdenlive
Geographic and scientific data visualization	QGIS、ParaView
Creating games and interactive works	Godot
Web browsing	Chromium

The uses listed are based on each application's official documentation. Executable files were checked in a single environment on October 1, 2026, Japan time. This is not an official guarantee that every dot receives the same software configuration.

This collection suggests a broad range of possible requests: create concepts for 3D assets or components, resize images for different uses, edit a short video, or turn geographic data into a map. Prototyping a game is one example among them. These possibilities follow from what the software is designed to do; we did not demonstrate each workflow in dots. A promising approach is to specify the deliverable's format and how it should be checked, then inspect the resulting files and request the next revision.

One question remains: Blender makes sense for 3D creation, but why include the Godot game engine as well? Godot is open source under the MIT license and supports creating and exporting projects from the command line. Beyond making assets, it offers the possibility of developing interactive works that respond to a user's actions. We have not found an explanation from OpenAI for choosing this configuration, however. This is an interpretation based on the software's uses. Godot licensing and official command-line guide

The configuration visible in the cloud environment examined

We also read information about the OS and computing resources from the same environment. The table below records the values visible there at the time of investigation, rather than specifications common to the product. Available resources and free space may change.

Item	Value observed
OS	Debian GNU/Linux 13 (trixie), detailed version 13.6
Kernel and architecture	Linux 6.18.44、x86_64
Reported CPU model	AMD EPYC 9V74 80-Core Processor
Reported CPU clock	Approximately 2.596 GHz
CPUs reported as available	9 logical CPUs; nproc and CPU affinity agreed
Reported virtual topology	1 socket, 9 cores, 9 threads
Memory visible to the environment	9.73 GiB total; no swap
Working storage	Overlay: 31.45 GiB total, with 29.65 GiB free when checked
Temporary storage	The /tmp tmpfs is 4.87 GiB; do not add it to the working storage as extra disk capacity

The 80-Core label in the CPU name does not mean this dot can use 80 cores. Nor does the virtual topology's nine-core figure guarantee nine physical or dedicated cores. The clock is a reported value at the time of observation, not a guarantee of sustained speed. cgroup information was unavailable, so additional limits on CPU time or memory remain unverified. These readings also do not establish the storage medium, persistence, or read and write performance.

Software availability, sufficient computing resources, and correct production of the intended deliverable each need to be checked in turn. For example, 3D rendering requirements, video processing time, CAD dimensions, and circuit correctness must be assessed for the actual job. This investigation shows one environment equipped with tools that could extend dots into a wide range of creative and analytical work. The official guide's advice still applies: check the files and results, not just the completed task status. Checking tasks and results

You decide what dots can access and do

What you can delegate to dots depends on the apps you connect and the actions you allow. Adding a communication channel does not by itself grant access to other apps or local files. Connection guide

Cloud research and tasks can continue while your computer is switched off. Access to local files or apps, however, requires permission for that device, which must be online with the ChatGPT app open. If a website requires a login, you may need to authenticate yourself. Describing dots as always available does not remove these conditions. Cloud and local device access

Before actions such as sending email or modifying files, Auto-review checks the instructions, permissions, safety conditions, and other relevant requirements. Ordinary reads remain subject to the connected service's permissions and limits, but do not require this additional review. Custom Rules let you set continuing boundaries, such as whether an action may proceed or requires confirmation first. They cannot override baseline safety restrictions. Actions and approvals guide and Auto-review scope

Even after delegating, you should check the content and sources of important results. Official material acknowledges that mistakes are possible and that a completed status does not necessarily mean the goal was achieved. Checking results

Stopping work also has several meanings. Pausing dots itself does not stop every delegated task or future scheduled run. Separate tasks must be checked in Activity, and recurring tasks in Scheduled. Stopping work does not undo changes already completed. How to stop work

Delegating Gmail and calendar work also means deciding how data is handled

The author has connected Gmail and a calendar to dots and asked it to act as an assistant.

This personal example establishes only that the connections and request have been made. The following discussion covers general considerations for that use; it does not describe the author's messages, appointments, task results, or impressions from using the service.

Reducing the need to supply relevant messages and appointments each time can make ongoing work easier to delegate. Email and calendar entries, however, may include information about senders, participants, and other third parties as well as the user. The convenience of giving dots more information must be considered alongside the broader data exposure and potential consequences of sending something to the wrong recipient.

Separate access to information from permission to act

dots uses the information and actions allowed by connected plugins. Asking it to read an email and draft a reply does not authorize sending that reply. Checking a calendar should likewise be distinguished from changing an event and notifying participants. The actions actually available depend on the connected service's permissions and settings. Connection guide and actions and approvals guide

A practical starting point is least privilege: allow only the connections and actions needed for the job. For example, require a review of the recipient and message before sending, or of the affected event and notification recipients before changing a calendar. These are operating suggestions, not settings the author claims to have configured. Custom Rules includes an example that prevents email sending. Instructions in a conversation are separate from the connected service's underlying permissions, so its actual permission screens should also be checked. Official controls guide and Custom Rules explanation

External emails and documents can also contain prompt injection: instructions unrelated to the user's request that try to induce actions such as sending information elsewhere. OpenAI describes a combination of model defenses, tool restrictions, pre-action checks, and safety monitoring. Auto-review checks details such as intended recipients and message content, but it is not the same as the user confirming every action; some actions can proceed under clearly scoped prior approval. Keep checks for important actions, and do not assume safeguards eliminate all mistakes. dots safety measures and limitations in the official FAQ

Disconnecting an app does not delete information already collected

According to the official FAQ, disconnecting a plugin stops new access through that connection. It does not remove information already brought into the dot's context. Deleting that context requires deleting the dot; there is currently no way to directly view, edit, or delete an individual dot's memories. Separately stored ChatGPT conversations and memories, generated files, and other data follow their own management and retention rules. dots data retention FAQ

Stopping a connection and dealing with retained information are therefore separate decisions. The general Privacy Policy also says data may be retained after deletion for safety, security, or legal reasons. Disconnection should not be treated as immediate, complete erasure. Privacy Policy retention conditions

Training use depends on the plan and the data's path

OpenAI says the use of conversations and tasks for training on personal plans follows the Improve the model for everyone setting. Business, Enterprise, and Edu data is not used for training by default. Proactive background research and its notes are not used directly for training, but information brought into eligible conversations or tasks may be used according to the applicable settings. dots training data explanation

For Gmail and calendars, ChatGPT's Google apps FAQ is also relevant. It says data fetched directly from connected apps and information derived from it generally are not used to train general-purpose models, while listing exceptions for conversations submitted as feedback, manually pasted or uploaded data, and data included in responses. It also says that, when training is turned off, app data appearing in conversations is not used to improve the models. Google app data usage conditions

That Google-specific explanation concerns connected apps in ChatGPT. Public documentation does not map it to every data path inside dots. It would be inaccurate to say either that all connected email is used for training or that it is never used through any path. Turning off training also does not disable safety review. The dots FAQ says limited human review, including for safety reasons, can occur even when training is off. dots privacy FAQ

Would an AI report an ethically troubling email?

Suppose an AI encounters ethically troubling correspondence while reading email. Would it independently report it to a third party or the police? This is a hypothetical question, not one based on a particular email or person. Ethical concerns, illegality, and imminent harm to a person are different things.

The public Model Spec says a model should not make enforcing laws or morality an independent objective, and should not engage in autonomous whistleblowing. That principle does not support describing an AI as independently reporting something externally solely because it considers it unethical. This is a general behavioral standard for models in OpenAI products, however, not a guarantee that deployed models follow it perfectly. Model Spec principles and scope

The AI model's own behavior must be distinguished from OpenAI's response as the service operator. The same Model Spec explains that potential usage policy violations are monitored automatically and that, after human review, cases involving imminent serious physical harm or other serious public safety risks may be referred to law enforcement. Model Spec explanation of operator responses

The Privacy Policy provides for disclosure to government authorities or third parties on grounds including legal obligations, addressing violations of terms, policies, or law, preventing fraud, and protecting safety. Separate guidance on law enforcement and other government requests describes disclosure under valid legal process, emergency disclosure to prevent danger of death or serious physical injury, and responses to lawful preservation requests. User notification also has exceptions, including legal prohibitions and emergencies. Privacy Policy and government data request guidelines

Certain illegal content is also subject to reporting policies. OpenAI publicly states that it reports child sexual abuse material (CSAM) identified on its services to the child protection organization NCMEC. This is an operator response; it does not mean the AI reports ethical concerns in general to outsiders. OpenAI's child safety policy

The dots-specific public material describes reading authorized information, reviewing actions before execution, and safety monitoring for harmful behavior during tasks. Those descriptions and the general Model Spec do not establish whether connected email is comprehensively scanned or which email contents might lead to operator review or reporting. A behavioral standard is not a confidentiality guarantee: neither 'it will never report anything' nor 'it reports ethical violations' is justified. Decide what information and actions to entrust to dots without speculating about the existence or scope of undisclosed monitoring. dots safety monitoring explanation

Eligible plans still face a gradual rollout and task usage limits

The official announcement dates the launch to September 29, 2026. The introduction guide, rechecked on October 1, describes a gradual rollout to the following groups. Even on an eligible plan, availability may take time to reach an account.

Plan	Currently stated conditions
Pro 100／200／500	Supported regions excluding the EEA, UK, and Switzerland
Business Premium	Rolling out across supported regions worldwide
Enterprise	Requires administrator enablement; disabled by default

Users under 18 are ineligible. Enterprise availability is described as a beta that includes Edu and Healthcare. Japan is not among the Pro exclusions above, but a subscription does not guarantee immediate access. Official help: eligibility and age requirements

A dot must be created in the desktop app or a desktop browser. After creation, it can also be used in mobile apps that have received the relevant update. Mobile web browsers are not supported. Official help: setup

The first dot is included with Pro and Business Premium at no additional charge. Official pricing lists Pro subscriptions at US$100, $200, and $500 per month. These are subscription prices, not the cost of a single request to dots. Official announcement and pricing guide

Usage limits require a separate distinction. The announcement and introduction guide say conversations with dots do not count toward ChatGPT usage limits. Work and Codex tasks started or managed by dots do count toward their respective normal allowances. Work and Codex share an allowance. Having the first dot included at no extra charge does not mean unlimited task execution. Usage limits and Work and Codex pricing

The communication guide confirms that voice calls with dots are available. The dots pricing material we reviewed refers to conversations but does not separately specify how voice calls are metered or capped. That wording alone does not establish that dots voice calls are free and unlimited. Work and Codex have separate pricing information for Desktop Voice, but we found no basis for applying it directly to calls with dots. dots usage limits and pricing including Desktop Voice

According to the announcement, plans include an allowance for substantial work. Work usage limits are expanded for the first month after the service's launch. No multiplier or specific additional amount is stated. Usage conditions in the official announcement

dots aims to carry work beyond a one-off research result, through updates and follow-ups as circumstances change. Users can delegate comparisons, document maintenance, and checks on ongoing work while continuing to supply constraints, review results, and make decisions. The idea is to delegate a recurring concern through the point where the information needed for a decision is ready. That is the kind of personal assistant suggested by the official scenarios.

Sources and scope

These primary sources were used for the September 30, 2026 draft and the October 1 additions and recheck. Venue comparisons, scheduling, and similar scenarios are OpenAI's illustrations. The author's personal example is limited to connecting Gmail and a calendar and requesting an assistant role. No email or calendar contents were investigated or quoted, and the reporting discussion is not based on a particular incident. Feature descriptions are distinguished from results in a specific environment, and model behavior guidelines from the operator's data handling policies.

1. OpenAI, Introducing dots — September 29 announcement, GPT-6 Astra, illustrative sales support workflow, pricing, and usage limits.

2. ChatGPT dots product page — Official examples of scheduling, document updates, and meal ordering.

3. ChatGPT Learn, Meet dots — Venue comparisons and ongoing work, eligible plans and regions, and gradual rollout.

4. Tasks and memory guide — Delegation, context, recurring tasks, and checking results.

5. Computers and apps guide — Cloud and local environments, taking over controls, Codex cloud environments, app permissions, and authentication.

6. Actions and approvals guide and safety explanation — Action checks, research permissions, and the limits of stopping work. The safety explanation also identifies Linux and Chrome in the cloud environment.

7. Official help: getting started, privacy and safety, and official pricing — Additional availability and pricing conditions.

8. Contacting dots — Confirms voice calling is available. This guide does not specify separate metering rules for dots calls.

9. dots privacy FAQ and Google app data controls FAQ — Connection permissions, retention and deletion, training use, and the scope and limits of safeguards.

10. Public Model Spec, August 18, 2026 edition — The principle against independently pursuing law or morality enforcement, and the distinction from operator monitoring and review. This is public behavioral guidance, not a detailed specification of how dots processes connected email.

11. Privacy Policy, Japanese version, and government data request guidelines — Retention, disclosure, and legal process. The versions reviewed were updated July 30, 2026 and effective January 1, 2026, respectively.

12. OpenAI child safety policy — Operator reporting policy for CSAM. It is not treated as evidence that all connected email is scanned.

13. Read-only investigation of the cloud environment (October 1, Japan time) — Checked executable files for 12 apps, Blender and Godot versions, and OS, CPU, memory, and working storage information. These are observations from the environment examined, not common product specifications or guarantees of dedicated resources or successful creative work.

14. Each application's official documentation, linked from the cloud environment table, plus the Godot FAQ and official command-line guide — Used to establish the software's purposes and capabilities, not to explain OpenAI's reasons for choosing it.