OpenAI recently introduced Dots , a new class of persistent AI agents designed to keep working after a user leaves a chat session. Powered by GPT-6 Astra, each Dot gets its own cloud computer and browser, can connect to more than 4,000 apps through OpenAI’s plugin ecosystem, and can continue pursuing assigned goals between conversations. Users can reach the same Dot through ChatGPT, Slack, Microsoft Teams and voice, with context carrying from one interaction to the next.

For business users, the interesting part is that OpenAI is trying to turn AI from something you consult into something you assign. A Dot can watch a project, check for changes, run background work, use connected software and return when a decision needs human input. OpenAI’s own guidance tells users to begin with work they want a Dot to “keep track of,” rather than treating each conversation as a fresh request.

But persistent AI is not a novel idea. Anthropic’s Claude Cowork is pushing agents deeper into desktop and knowledge work, while Meta’s Muse is built around completing multi-step tasks from its own cloud environment. Other agentic systems are moving in the same direction, from research and coding to browser use and workflow automation.

The competitive landscape is shifting from which model can generate the best answer to which company can build an agent people will trust with an ongoing job. For OpenAI, Dots are an attempt to turn the company’s enormous user base into a persistent work layer that stays involved after the conversation ends.

What Makes A Dot Different?

The simplest way to understand Dots is to stop thinking about a typical chat-based operation or even a longer-form project or coding task. A normal LLM interaction has a natural stopping point. Ask for a market analysis, receive the analysis, perhaps refine it, then move on. Instead, consider agentic Dots to be given an ongoing responsibility and operate in a mostly-autonomous mode.

A classic chatbot is reactive. You ask, it answers. But a persistent agent is expected to notice, remember and act. Think of the difference between asking an intern for a report and telling that intern to own the report from now on. That is the product OpenAI is trying to build.

OpenAI gives examples such as keeping sales proposals current, processing new interviews into content, following software feedback and maintaining launch materials as a product changes. Dots can initiate cloud work, create recurring tasks and continue working when the user’s own computer is off. They can run several tasks in parallel and send updates when something requires attention.

Dots must initially be created from the ChatGPT desktop app or from ChatGPT in a desktop web browser. They cannot currently be created through mobile, and mobile web is not supported. Users who want a Dot to work with files, software or other resources on their own computer must connect that machine through the ChatGPT desktop app.

The Real Advantage Is Persistence, Not Smarter Answers

The frontier AI model markets are shifting to a point where model capability by itself is becoming a less useful way to distinguish AI products. OpenAI, Anthropic, Meta and others now offer systems capable of research, coding, computer interaction and multi-step work. The competitive dynamic is moving toward how those capabilities are packaged into something people can actually delegate work to.

One branch of agentic AI is being built around software that works inside systems a company already owns or controls, whether that means an employee’s computer, an enterprise application or a private cloud environment. OpenAI is taking a different approach with Dots.

Each Dot runs on separate OpenAI managed computing infrastructure, giving the agent its own persistent environment rather than depending entirely on the user’s machine. That can make long running work easier, but it also shifts more of the agent’s execution, data handling and operational control outside the customer’s own infrastructure.

Each agent has a cloud computer that retains its state. It can keep browser sessions, files and software available between periods of work. Users can separately connect a personal computer if a task requires local files or applications, though that machine has to remain online and running for the ChatGPT desktop app to operate with local work.

One Japanese developer recently demonstrated this type of workflow by asking a Dot to inspect folders on a connected Mac. The Dot delegated the work to Codex, executed a read-only task and returned the correct result. The author saw the key change as a shift from a person opening Codex and directing it to a Dot deciding when to invoke Codex on the person’s behalf.

Another practitioner describes handing Dots roughly eight hours of overnight work, with the agent coordinating Claude Code and Codex on a Mac and reporting through Slack or a phone.

Where Dots Could Actually Be Useful

The strongest Dot use cases are likely to be jobs that are repetitive without being completely mechanical. A scheduled script can already send a report every Monday, for example, but scheduled tasks have always been possible.

More useful are continuous monitoring and operational tasks. OpenAI explicitly positions Dots for research, data analysis, document preparation and software work. For example, a product manager could ask a Dot to monitor customer feedback, group recurring complaints and prepare a summary when a pattern changes. A sales executive could have one keep a proposal current as pricing, product information or customer requests change. A researcher could assign one to watch for new information and rerun analysis when relevant data appears.

How Dots Compare With Claude Cowork And Other Agents

Anthropic’s Claude Cowork attacks this area of semi-autonomous agentic operation by by letting users describe an outcome and allowing Claude to plan and carry out the steps. Greenlit Books, which compared the two products using vendor documentation, characterizes the difference by saying that Dots are built around an ongoing agent that keeps responsibility moving between conversations, while Cowork is oriented toward completing a described knowledge-work outcome.

Meta has moved in the same direction with Muse, an agent that can perform tasks such as sending emails, completing forms and handling travel related work from its own cloud environment.

Agentic AI is quickly becoming less of a category and more of a continuum. Dots, Claude Cowork and Meta Muse are converging with more fully agentic harnesses such as Hermes and OpenClaw, while edge based systems push autonomy onto local machines, browsers and enterprise infrastructure. The real distinction is no longer what these products are called, but how much autonomy they have, where they run and who controls the environment.

Part of the reason why companies are starting to settle into these agentic system is that it helps to build a competitive moat. Specifically, switching cost.

Moving from one model or LLM to another is relatively painless. Moving away from an agent that has learned how your organization works, accumulated context, built history and been connected to your business software could be much harder.

The Cost Question Is More Complicated Than It Looks

One issue regarding always-on and always-operating agents is their cost. While Dots aren’t priced separately, OpenAI says eligible Pro users can access Dots as part of Pro 100, Pro 200 and Pro 500 plans, subject to region and rollout restrictions. Its current Pro plans cost $100, $200 and $500 per month respectively. Business Premium and Enterprise access are rolling out separately.

OpenAI says conversations with a Dot do not count against regular ChatGPT usage limits. Tasks a Dot launches in products such as Work or Codex still count against those products’ limits, and each plan includes an allowance for what OpenAI calls “deeper work.” OpenAI has not publicly specified the size of that allowance in its documentation.

That makes cost forecasting less straightforward. According to OrcaRouter , a Dot is closer to renting a packaged worker, whereas API models are metered components whose token costs can be tracked directly.

Another big challenge with always-on, semi-autonomous Dots is that they need access to data and systems to work effectively. This presents security, privacy, compliance and governance challenges.

OpenAI allows users to connect services such as Gmail, Google Drive and GitHub through plugins. Existing permissions carry through to the Dot, and users can restrict actions. One connection might permit email reading without permitting the sending of email. Connecting a messaging service such as Slack does not automatically give the Dot access to other applications or a local computer.

OpenAI has built approval mechanisms around sensitive actions. Before an action affects an account or shares information, an automated review checks permissions, user instructions and safety requirements. Some actions can proceed automatically, some require approval and some must be handed back to the user. OpenAI gives changing a password as an example of an action the human must perform.

Users can also create custom rules requiring approval before sending customer messages or handing file deletion back to a person. But OpenAI warns that these rules remain instructions that the system can make mistakes in following.

This is an issue for Dot systems that have access to your documents, email, and enterprise systems such as CRM and ERP.

To solve enterprise compliance and governance needs, Enterprise Dots are disabled by default and must be activated by an administrator. The company provides controls for local computer access, plugin permissions and supported audit records through its Compliance and Analytics APIs.

There is another drawback that has less to do with technology. Agents need instructions. Not prompts in the old sense, but operating rules and processes. A person must still define what the agent is responsible for, what a good result looks like, which sources matter, when it should interrupt someone and what decisions require approval.

The problem is that processes often rely on unwritten habits, something people call “tribal knowledge”. Two departments may define the same metric differently. Approvals may happen informally in Slack. A spreadsheet may be regarded as the official source until somebody remembers that another spreadsheet is newer.

Humans compensate for that ambiguity every day, but persistent agents will have a harder time.

So, Do You Need An OpenAI Dot?

If you mostly use AI systems for simple tasks or direct response activities for writing, research, brainstorming, analysis or isolated questions, you may see little benefit from giving an AI its own persistent computer.

Dots become compelling when you have recurring work that keeps coming back after a single task or project need. These persistent agentic systems provide the most benefit when there are tasks you repeatedly check and perform.

To determine if Dots are a fit for you, ask which project needs someone to remember its state. Ask what work could continue without you watching it.

Ask whether you can describe the desired result clearly enough that another person could take ownership of it. If you can, Dots may represent a more meaningful shift than another round of smarter models. If you can’t, an always-on, semi-autonomous AI agent probably will not fix the problem.