OpenAI Dots Turns ChatGPT Into an Always-On AI Agent

OpenAI has spent years improving ChatGPT's ability to answer questions, reason, code, and complete complex tasks. OpenAI Dots change the interaction model.

Introduced on September 29, 2026, Dots are what OpenAI calls “always-on agents.” They are powered by GPT-6 Astra, run on their own cloud computers, can use their own browser, connect to more than 4,000 applications through OpenAI's plugin ecosystem, and continue working toward a goal between conversations. OpenAI's Dots announcement

The key difference isn't simply that Dots can automate tasks. AI agents already do that. Dots are designed to take ongoing responsibility.

A traditional assistant waits for a request. A task agent may execute a multi-step instruction. A Dot can remain attached to a goal, watch what changes, continue working, use other tools and agents, and return when it needs a decision.

That makes Dots less like another ChatGPT mode and more like OpenAI's attempt to build a persistent AI layer between a user and the software they use.

What are OpenAI Dots?

A Dot is a persistent AI agent inside the ChatGPT ecosystem.

According to OpenAI, users give a Dot goals and define how much it can do independently. The agent can then work through those goals, adapt as circumstances change, and keep progressing even when the user is no longer actively interacting with it.

OpenAI's examples show what that means in practice. A developer's Dot could watch customer feedback, identify recurring requests and bugs, build and test fixes and return completed pull requests for review. A scientist's Dot could rerun analysis as new data arrives. A product team's Dot could update launch materials when requirements change. A sales agent could keep an enterprise proposal aligned as technical requirements and test results evolve.

OpenAI also says one early tester's Dot noticed that he had forgotten to invoice a publication, prepared the invoice and sent it after receiving approval.

The shift is therefore:

Prompt → Task → Responsibility

Instead of telling the AI what to do every time, the user increasingly defines what outcome it should remain responsible for.

The features that make Dots different

The foundation of each Dot is its own cloud computer. It can browse the web, analyze information, create files and run tools inside an isolated environment. Work can continue there while the user is doing something else.

Users can inspect the Dot's computer and follow ongoing work. They can also optionally connect their own computer, giving the Dot access to approved local files and tools. OpenAI says sandboxing separates each Dot's workspace and prevents it from modifying the safety systems responsible for controlling its behavior.

Dots also have persistent context. They can learn a user's preferences, goals and standards over time and carry context across conversations and communication channels. OpenAI says a Dot can be reached through ChatGPT, Slack and Microsoft Teams, with voice interaction available and text messaging planned.

Another important feature is proactive research. When a Dot is not actively working with the user, it can examine information in permitted connected sources to identify developments that might matter. OpenAI gives the example of detecting a change to someone's travel plans.

That background access is intentionally restricted. Proactive research uses read-only tools, meaning the background process cannot directly send messages, change connected-app content or control the user's browser or computer. If the Dot decides action is required, the next step must pass the usual permissions and safety checks.

The result is an architecture closer to:

Observe → Understand → Decide → Request authority where necessary → Act → Continue

rather than simply responding to prompts.

Dots can delegate work to other OpenAI systems

Dots are also important because they sit inside a much larger OpenAI stack.

OpenAI already has ChatGPT Work for substantial delegated work and Codex for software engineering. Its recently introduced Agents API gives developers infrastructure for long-running cloud agents that can maintain context, use tools, work with files and coordinate subagents over extended periods. OpenAI's Agents API announcement

Dots can coordinate tasks using these systems rather than trying to do everything in one model interaction.

That creates a hierarchy that increasingly looks like:

User → Dot → specialized agents, apps and tools → completed outcome

This may be one reason OpenAI launched Dots now. The company already had many of the individual components required for persistent agents. GPT-6 Astra supplies the reasoning and computer-use capability. ChatGPT provides the user interface and context. Plugins connect external systems. Codex handles coding. Work executes larger projects. Dots ties these capabilities together around a persistent objective.

OpenAI's own product direction suggests the company is moving beyond building a chatbot toward building infrastructure through which AI can execute work across other software.

Why launch Dots now?

There is also a larger market reason.

Agentic AI is moving from experimentation into enterprise planning. Gartner's 2026 survey found that only 17% of organizations had deployed AI agents, but more than 60% expected to deploy them within two years. Gartner cautions that fully autonomous agents are still unsuitable for most enterprise use cases, but the adoption intent is unusually aggressive. Gartner's 2026 agentic AI analysis

OpenAI therefore has an incentive to establish ChatGPT as the layer through which those agents are created and controlled before organizations standardize on competing platforms.

The timing also follows Meta's September launch of Muse, giving OpenAI a direct competitor in an emerging category where the product is no longer simply a model or chatbot. It is an agent with memory, tools, permissions and the ability to act.

Which markets could Dots affect?

The immediate impact is on knowledge work.

Software development, sales, research, marketing, customer support, finance, procurement and administrative work all involve recurring tasks spread across multiple systems. Dots is designed specifically for environments where work changes over time and cannot be reduced to a single fixed automation.

OpenAI is already moving beyond personal Dots into specialist Dots for organizations. These agents can receive their own identity, credentials and access to specific business systems. OpenAI says it has tested the approach internally across procurement, invoice processing, email marketing, customer support and commercial contracting.

The company is also working with Microsoft to integrate specialist Dots with Microsoft Agent 365, giving organizations enterprise governance and security controls around these agents.

But the larger impact may eventually fall on the SaaS market itself.

Gartner estimates that as much as $234 billion of enterprise application spending could be exposed to what it calls “agentic arbitrage” by 2030, representing roughly 20% of enterprise SaaS spending. The argument is straightforward: if an agent can operate across several systems and produce the outcome directly, employees may spend less time inside each software interface. Gartner's research on agentic AI and SaaS spending

Consider an employee asking a Dot to analyze an important customer, identify churn risk, prepare a proposal, schedule a meeting and update internal records. The Dot could potentially interact with CRM, analytics, email, calendar and document systems without requiring the employee to manually move through each application.

The applications do not necessarily disappear.

The interface becomes less important.

That could shift software competition toward APIs, data access, permissions, context, reliability and compatibility with agents rather than dashboards and human seats alone.

Dots also makes the AI safety problem more real

The same capabilities that make Dots valuable increase the consequences of mistakes.

A chatbot producing a bad answer creates one type of problem. An agent capable of accessing email, files, browsers and business systems can turn a reasoning error into an external action.

OpenAI therefore built several control layers around Dots. Users choose which applications the agent can access. Custom Rules can specify which actions are allowed, which require approval and which should be blocked.

Before consequential actions such as sending email or modifying files, a separate safety system called Auto-review evaluates the planned action against the user's instructions, permissions and safety requirements. Certain actions always remain with the human. OpenAI says changing passwords or transferring money between financial accounts cannot simply be delegated to a Dot, while purchases require approval.

That architecture says something important about where agentic AI is heading.

Model intelligence is no longer the only problem. Identity, access, authorization, monitoring and containment are becoming core AI product features.

Dario Amodei's warning is becoming easier to understand

Anthropic CEO Dario Amodei recently described this tension unusually clearly.

In his September essay, “We Must Pace the Frontier,” Amodei argued that AI capabilities are beginning to advance quickly enough that safety practices may struggle to keep pace. He specifically cited the risks of systems being misused for cyberattacks and bioterrorism, economic disruption and, at the extreme, humans losing control of increasingly autonomous systems. His argument is not to stop AI development altogether, but to slow capability advancement where necessary so alignment, safeguards and independent evaluation have time to catch up. Dario Amodei's “We Must Pace the Frontier” essay

The same Amodei is also extremely optimistic about AI's potential. He has argued that advanced AI could help cure most major diseases within the next five to ten years, accelerate economic growth and produce major advances in biology and medicine.

The five-to-ten-year disease timeline is a prediction, not an established medical forecast, and some researchers have challenged how realistic it is because biological discovery still has to pass laboratory validation, clinical trials and regulation. But the broader tension in Amodei's argument is already visible in products such as Dots.

AI is becoming more useful because it can do more without us. The same capability also gives mistakes a larger blast radius.

An AI that only writes text can misinform someone. An AI that can operate computers, coordinate agents, access systems and continue working autonomously can potentially affect those systems.

Dots makes that debate practical rather than theoretical.

Is Meta Muse a rival to OpenAI Dots?

Yes.

Meta launched Muse on September 8, three weeks before OpenAI introduced Dots. Meta describes Muse as a personal AI agent that does more than answer questions. It runs inside a dedicated Muse Secure VM, has its own browser, works across connected applications, learns from conversations and can independently advance long-term goals. Meta's Muse announcement

That makes Muse and Dots competitors at the category level.

Both are trying to move AI from a reactive assistant into a persistent agent with memory, access to tools and the ability to continue working after the user stops actively interacting.

Their positioning, however, is different.

Meta's early positioning is more aggressively focused on the consumer layer. Muse works through WhatsApp and Meta is bringing the agent to its AI glasses, allowing Muse to respond to what the wearer is looking at and continue work in the background.

Meta has also begun expanding Muse into businesses through the Meta Enterprise Platform and Muse for Small Business, so it would be wrong to treat Muse as permanently consumer-only.

OpenAI's initial strength is different. Dots is tightly connected to ChatGPT, Codex, ChatGPT Work, enterprise applications, Slack, Teams and thousands of plugins. Its specialist Dots strategy is explicitly designed around defined organizational responsibilities.

The competitive positioning is therefore becoming clear.

Meta approaches the personal agent through social communication, devices, commerce and consumer distribution. OpenAI approaches it through ChatGPT, professional work, developers, enterprise systems and specialized agents.

Both ultimately want to own the interface between the user and the digital world.

Who can access OpenAI Dots now?

Dots began rolling out on September 29, 2026, and access currently depends on both subscription and region.

For individual users, Dots is rolling out to ChatGPT Pro subscribers outside the European Economic Area, Switzerland and the United Kingdom. Those regions are excluded from the initial personal Pro rollout.

For businesses, ChatGPT Business Premium users can receive Dots across supported ChatGPT markets. OpenAI's Business Premium seat currently costs $100 per user per month on annual billing or $125 on monthly billing. Standard Business seats do not receive Dots in the initial rollout.

Enterprise users, including eligible Edu and Healthcare workspaces, can try Dots through a beta when their workspace administrator enables it.

That means Free, Go, Plus and standard Business users are not part of the initial Dots rollout.

The first Dot is included with eligible Pro and Business Premium plans at no additional cost. OpenAI says conversations with a Dot do not count against normal ChatGPT usage limits, while tasks a Dot starts or manages through Codex or ChatGPT Work continue to consume those products' allowances.

Initial setup must be completed through the ChatGPT desktop app or desktop browser. After setup, users can communicate with the Dot through the mobile ChatGPT app as well. Rollout is gradual, so an eligible account may not receive access immediately.

From asking AI to giving AI responsibility

The most important thing about OpenAI Dots is not another feature list.

It is the change in relationship between humans and AI.

Chatbots were built around questions.

The first generation of agents was built around tasks.

Dots is being built around responsibility.

“Check my customer feedback” becomes “keep track of what customers are struggling with and bring me fixes.”

“Update this proposal” becomes “keep this proposal aligned as the customer changes its requirements.”

“Analyze this data” becomes “keep this research current as new evidence arrives.”

That shift creates a much larger opportunity for OpenAI than simply winning another chatbot benchmark. If Dots becomes the layer through which users communicate with applications, delegate work, coordinate specialist agents and manage ongoing objectives, OpenAI can sit between the user and a substantial part of the software economy.

Meta Muse suggests OpenAI will not have that market to itself.

But Dots shows that the competition is moving beyond who can build the smartest chatbot.

The next contest is increasingly about which AI company can build an agent capable enough to take responsibility, connected enough to get useful work done and trustworthy enough that users are willing to let it act on their behalf.

That is a much bigger test than answering a prompt.