OpenAI Dots are agents designed to keep working across conversations. Announced on September 29, 2026, they use GPT-6 Astra, have a cloud computer and browser, and retain context for ongoing projects. OpenAI says its plugin ecosystem connects them to more than 4,000 apps.
The practical question is how to delegate a useful piece of work while keeping its scope clear. This guide covers access, permissions, and a suggested first experiment. It is based on official documentation; we have not independently tested Dots.
What can OpenAI Dots do?
According to the launch announcement, you can reach your dot through ChatGPT, Slack, and Teams, including voice interaction. You can inspect its cloud computer and optionally connect your own device. OpenAI illustrates workflows involving tested code fixes, revised launch assets, updated analysis, and content drafts.
These are product examples, not evidence that every workflow will succeed without intervention. The announcement also previews specialist dots for dedicated organizational responsibilities through focused enterprise pilots.
Who can access OpenAI Dots?
The getting-started guide describes a gradual rollout. Access can take several days to appear. Its plan and regional conditions are:
| Plan | Documented launch availability |
|---|---|
| Pro | Markets excluding the European Economic Area, Switzerland, and the UK |
| Business Premium | Supported ChatGPT regions |
| Enterprise | Beta when enabled by the workspace administrator; initially off by default |
| Free and Plus | Not listed among the initial rollout plans |
Create a dot in the desktop app or desktop browser. The guide says mobile creation and mobile web are unsupported; mobile-app conversations follow setup when available. It also says dots cannot initiate calls to you at launch.
Check your account’s available features before purchasing a plan specifically for Dots. A rollout announcement does not establish that a particular account already has access.
Is Dots included, and is usage unlimited?
The announcement says the first dot is included in Pro or Business Premium, with a deeper-work allowance and expanded launch limits. It distinguishes dot conversations from tasks started in Codex or ChatGPT Work, which use their usual limits.
The getting-started guide separately says dots usage will not count toward eligible plan allowances for the next month, with later terms to follow. These pages use different wording. Check the terms shown for your account; do not interpret the launch offer as permanent unlimited execution.
Background research and authorized tasks are different
The privacy and safety FAQ distinguishes proactive research from actions. Proactive research can read permitted sources and retain notes, but its tools cannot directly send messages, change plugin content, or control a browser or computer. Follow-up actions and previously authorized background tasks remain subject to the usual rules.
Custom Rules can set boundaries for supported actions. They cannot disable mandatory safeguards. Some sensitive operations require you to take over. Activity View lets you inspect progress and redirect work.
For a first project, this distinction suggests a useful scope: ask the agent to collect evidence and prepare a draft, then evaluate that draft before expanding what it may change.
Memory and connected apps: what to understand
The same FAQ says dot context and ChatGPT Memory can share information. Disconnecting an app stops new access but does not erase information already retained. Individual dot memories currently cannot be directly viewed, edited, or deleted; deleting the dot removes its own context, while separately stored files and ChatGPT memories have their own controls.
Our practical suggestion is to begin with a small project folder containing only the material needed for the experiment. Broader access can wait until you have a reason for it.
How the cloud workspace is protected
OpenAI’s security explanation describes isolated cloud workspaces, sandbox restrictions, and a Linux environment with Chrome. Supported sign-in flows keep passwords outside the model’s context. Action checks and monitoring add protection against mistakes and malicious instructions encountered in external content. OpenAI also acknowledges that these measures do not eliminate errors.
For business, Enterprise, and Edu workspaces, content is not used for model training by default. Personal-plan users control eligible use through the model-improvement setting. Review those controls before connecting work accounts.
A first OpenAI Dots experiment for a small website
Here is an editorial test you can adapt if your account has access. It is a proposed workflow, not a reported product benchmark.
Create a small set of ten fictional reader comments about a website. Include repeated requests, one unclear complaint, and one request that conflicts with your editorial policy. Give the dot the policy and ask for a review document.
- Goal: identify recurring reader needs and propose three improvements.
- Sources: use only the supplied comments and editorial policy.
- Evidence: link each proposal to the relevant comment identifiers.
- Deliverable: prepare a table with the issue, supporting evidence, proposed change, and unresolved questions.
- Boundary: prepare drafts only; require review before publishing or contacting anyone.
After the first result, add two new comments that change the priority of one issue. Ask for a revised report that explains what changed. This tests whether the agent incorporates new evidence while preserving instructions.
| Check | What a useful result looks like |
|---|---|
| Traceability | Every proposed change has supporting comment IDs |
| Accuracy | No invented reader statements or unsupported statistics |
| Revision quality | New evidence changes the relevant recommendation |
| Scope control | The work stays within the supplied project and permissions |
| Review effort | You can verify the report without rebuilding the analysis |
Keep a record of the time you spend preparing the task and reviewing the result. An agent that produces more material is not automatically saving work. The useful comparison is the total effort needed to reach an acceptable result.
Write a goal with a clear definition of done
A request such as “improve my website” leaves too much open. A more useful brief identifies the pages or files, the problem to investigate, the evidence to use, and the output you expect. Add a stopping condition, such as delivering three documented proposals or pausing when the source material is contradictory.
For recurring work, specify the schedule, sources, and circumstances that require your input. Start with one responsibility. Expand it after you understand the recurring errors and the effort needed to review the results.
Our AI tool testing checklist provides a broader evaluation framework. If you want to understand persistent context as a separate technical concept, read our Hindsight agent memory guide.
OpenAI Dots makes ongoing delegation the central product idea. Its usefulness for your projects will depend on the available integrations, clear instructions, access limits, and the quality of completed work. A small, reviewable experiment is a practical way to assess that fit.
Official sources
- Introducing dots — OpenAI, September 29, 2026
- Getting started with your dot
- Dots privacy, security, and safety FAQs
- How OpenAI builds safety, security, and privacy into dots
Reviewed September 30, 2026. Availability and usage terms may change during rollout.