What Is OpenAI Dots? Agents, Access and Limits
What is OpenAI Dots? Understand the agent, its launch access, cloud work and limits, then decide which ongoing responsibility is worth handing over.
OpenAI dots is an agent for ongoing work inside ChatGPT. You give it a responsibility, then return with corrections or new priorities while the work continues. It is not the name of a new model you select for an ordinary chat. OpenAI’s dots overview names GPT-6 Astra as the model behind it. The useful question is what work you would actually want it to keep moving.
This guide covers the distinction, who can get access, and how to judge a first assignment. If you already have dots and want a practical walkthrough, use how to use OpenAI dots. Product details below were checked against OpenAI’s public documentation on 30 September 2026. This is an explanation of those documents, not a hands-on review or a claim about measured time savings.
Dots, ChatGPT and the model are different things
Think about a supplier review. A chat request might be: “Summarise these notes.” You read the answer, edit it and move on. An ongoing responsibility is broader: keep track of the review, notice which questions remain unanswered, and prepare the next draft when new information arrives. That second shape of work is the reason to consider an agent.
The distinction is about the responsibility you assign, not whether a reply sounds clever. A polished paragraph can still miss a deadline, misunderstand who owns an action, or turn a tentative idea into a commitment. For the underlying vocabulary, AI agent vs LLM explains why a model and the system using it should not be treated as interchangeable.
It is also worth being precise about Muse. Coursium’s Muse and Muse Spark explainer separates Meta’s agent from its model. Apply the same habit here. Before comparing two products, establish whether you are comparing models, chat interfaces or agents that can act. A ranking of model answers does not settle which agent fits your accounts or your review process.
Who can use dots at launch?
OpenAI’s launch access guidance says Pro rollout excludes the EEA, Switzerland and UK. Business Premium access covers supported ChatGPT regions. Enterprise beta access needs an administrator to enable it. Rollout is gradual, so an eligible subscription does not guarantee the feature has reached your account today.
Check your actual plan, location and workspace before buying anything for this feature. A general ChatGPT subscription page and a dots access page answer different questions. If the option is missing, first check the rollout guidance and which workspace is selected. Do not assume that paying for a different tier removes a regional restriction.
The launch announcement says the first dot is included in Pro or Business Premium, with an allowance for deeper work. Tasks in ChatGPT Work or Codex still use their usual limits. Treat inclusion as a plan feature, not a promise of unlimited work. Check current account terms before planning a recurring workload around it.
Where the work happens
OpenAI’s computer and app guide distinguishes the dot’s cloud computer from an optional connection to your own computer. Cloud work can continue with your device off. Local work needs the connected computer online with the ChatGPT app open. The location matters when the task depends on a file or application on your machine.
For example, public research for a supplier comparison has a different requirement from editing a workbook stored only on your laptop. Write that requirement into the brief. Name the file and its location rather than assuming the agent will find the newest version. If you move a source or change its name, tell it. Otherwise you can end up reviewing an accurate answer to an outdated document.
Keep the first assignment easy to inspect. A comparison with source links and a clear list of missing information is easier to check than an instruction to “handle procurement”. The broader request hides many decisions: which suppliers count, who may see the prices, and whether an email should be sent. Those decisions belong in the task before you grant more access.
What an ongoing brief looks like
Here is a proposed exercise, not an observed dots result. Suppose you manage a small team and need to decide whether to keep the current meeting venue. You have the old agreement, a new quote and notes about what the team disliked. The aim is a decision brief. It is not a booking.
Compare the current venue with the new quote using only the documents I provide. Organise the draft by capacity, travel, included services and cancellation terms. Cite the document section behind each claim. Mark missing details as unknown. Keep an unanswered-questions list as I send updates. Do not contact a venue, make a booking or share the documents. Bring the comparison back for my review.
This names a result, sources, a structure and boundaries. It also defines what new information should change. If the quote leaves out equipment, the correct output is a question about equipment, not an invented assumption that it is included. You can improve the brief using the same method in writing a prompt that works first try.
On your next visit, give a concrete update: “The new venue includes the projector but charges separately for microphones. Update the comparison and keep the other unknowns.” Then check that the change appears in the right place. This is more useful than praising the draft and hoping the agent understood the correction.
Follow-up is useful only when its scope is clear
According to OpenAI’s tasks and memory guide, assigned work can continue between conversations, and fixed recurring work needs a saved schedule. Ask for confirmation of what was saved. A request to keep a project moving and a request to send a report every weekday are different instructions.
For a recurring venue review, define the trigger: a revised quote or a changed headcount. Define the update: revise the comparison and show what changed. Define the audience: you, rather than the whole team. Define the stopping point: when the venue decision is made. Without those details, routine updates can become noise and an old task can outlive its purpose.
Pick a responsibility with a clear owner. If two people give conflicting budgets, tell the agent whose decision controls the draft. Do not ask it to resolve a disagreement it cannot evaluate. A brief should make unresolved decisions visible so the people responsible can settle them.
What to check in every result
Start with the source, not the confidence of the wording. Check the quote’s date, the named venue and the cancellation clause. Look for claims that have no supporting passage. Ask whether the document described an estimate or a firm offer. Those differences often matter more than how neatly the comparison is written.
Then check what actually happened. A completed draft and a sent message are different outcomes. A note saying a task finished is not proof that a file was saved in the correct place. Open the deliverable and inspect it. The routine in checking an AI answer when you are not the expert helps when the subject is unfamiliar.
- Does the result answer the decision you named?
- Can you trace important claims to current sources?
- Did it preserve unknowns and disagreements?
- Is the deliverable available where you expected it?
- Did any action go beyond the brief?
Permissions do not replace judgement
OpenAI’s safety explanation describes safeguards and separate action checks, while acknowledging that dots can make mistakes. Treat those protections as part of the system. They are not a substitute for reading consequential work or deciding whether an account should be connected.
If you want a boundary that applies beyond one task, OpenAI dots permissions and controls walks through the distinctions between app access, instructions and memory. The most useful starting point is modest: a draft you can inspect, from sources you can share, with no external action needed.
What to do this week
- Check official access guidance and the account where you expect to use dots.
- Choose an ongoing responsibility with a visible result and a clear end.
- Write down the sources, allowed actions, audience and review point.
- Run a small draft task and check it against the original material.
- Add recurring work only after you understand the first result.
The durable skill is deciding what to delegate and how to verify it. Coursium teaches practical AI skills for work through lessons and practice tasks on your phone. Start with a responsibility you understand well enough to check, and widen it only when the evidence from your own work supports that decision.