OpenAI Unveils "Dots," Always-On AI Agents That Never Clock Out — as Its Next Model Is Shelved Over Safety Fears

OpenAI CEO Sam Altman unveiled "Dots" — always-on AI agents that work in the background across ChatGPT, Slack, Teams and voice — at the DevDay 2026 keynote in San Francisco on September 29, 2026. (File photo: Altman at TechCrunch Disrupt San Francisco 2019.) Photo: Steve Jennings/Getty Images for TechCrunch via Wikimedia Commons, CC BY 2.0.

For weeks, the rumor had a one-letter name: "o." Code sleuths found it buried in ChatGPT's configuration files — an always-on assistant with its own email suffix, spotted on a Pro upgrade page and then scrubbed. On Tuesday afternoon in San Francisco, Sam Altman finally showed the world what it became.

Its name is Dots. And it never logs off.

At OpenAI's DevDay 2026 keynote — the company's biggest yet, with what OpenAI called more than 20 major announcements — Altman introduced Dots as "remarkably capable, always-on agents built to handle everything." Not a chatbot you prompt. A colleague that keeps working after you close the tab: monitoring customer feedback and shipping bug fixes, rerunning a scientist's analysis as new data arrives, managing your calendar while you sleep. Powered by GPT-6 Astra. Available starting Tuesday.

It is OpenAI's boldest move yet in the war for the agent era — and it landed less than 24 hours after the company killed its next flagship model over safety fears.

What Dots actually are

The conceptual leap is simple and total. ChatGPT, for all its power, is reactive: it waits for you to ask. Dots are persistent. You give a dot a goal and ground rules, and it keeps working in the background — proactively updating you at milestones, pinging you when it needs a human decision, and otherwise staying out of your way.

"Today, you can start with your primary dot, give it a name, and make it your own," OpenAI said in its announcement. "Over time, we envision teams of Dots working together on your behalf."

The naming part is deliberate. Like Meta's Muse assistant, Dots get their own identities. "By giving Dots their own identities, they become an extension of our team," said Holly Li, a member of OpenAI's product team, during the keynote. Each dot gets its own browser, works across external apps — calendars, Slack and more — and can be reached wherever you already are: the ChatGPT app, text messages, Slack, Microsoft Teams and other workplace platforms, or by voice. Text-message support is coming soon, according to TechCrunch.

OpenAI also sketched "specialist Dots" — agents provisioned with specific identities, credentials and tools through a company's existing systems, so a finance dot and an engineering dot operate with different permissions. The company said it is already working with Microsoft to integrate Dots into Microsoft's Agent 365 security controls, an early acknowledgment that always-on agents with real credentials are a security problem as much as a product.

Available today — if you pay

Dots are live now for ChatGPT Pro and Business Premium subscribers in eligible markets, launchable from either ChatGPT or Codex. Enterprise, education and healthcare customers can try a beta once their workspace admin enables it — and notably, it ships off by default there, a quiet nod to the compliance headaches of deploying autonomous agents inside regulated organizations.

The tiering matters. Persistent agents that burn compute around the clock are expensive to run, and OpenAI is clearly using Dots to draw a brighter line between its free and paid universes. The leaked "o" references had already suggested the always-on assistant would live behind the Pro paywall; DevDay confirmed the shape of that bet. The most capable form of AI OpenAI has ever shipped to consumers is, for now, a premium product.

ChatGPT Space: the shared desk

Dots did not arrive alone. OpenAI also unveiled ChatGPT Space, a shared environment where teams work together alongside Dots — which, from Altman's onstage demo, resembles a shared Google Doc with agents sitting at the table.

The framing is the tell. OpenAI's DevDay recap described opening up ChatGPT "as a shared surface where humans and agents can collaborate and where developers can directly launch new native experiences to our collective 1.2 billion weekly users." ChatGPT is no longer an app; it is becoming a platform — a place where third-party developers build plugin extensions with their own sidebar homes and interactive panels, and where agents from multiple vendors could eventually coexist.

Developers got plenty beyond the headline act: Codex can now run anywhere — on a computer, remotely from a phone, or in the cloud from any device — with reusable development environments that give teams shared setups with approved settings and permissions.

The model that didn't make it

All of this unfolded under a cloud OpenAI chose not to hide. On Monday — one day before DevDay — the company announced it was scrapping the public release of GPT-6.1 Astra, which had been expected within days or weeks with a potential October debut. Its own researchers had raised safety concerns; the model, the company said, had not met its safety and alignment bar.

"For anything regarding safety and alignment, there's a trade off. You really do need to find what's the right line between staying within scope, but also avoiding laziness in terms of how the model actually pursues tasks even when it hits friction," Saachi Jain, OpenAI's head of safety systems, said in an emailed statement to Barron's on Tuesday. The shelved model, she added, "didn't quite meet the bar" in terms of staying within scope and authorization and how it communicates back to the user.

The company said other new models that do meet the bar are coming soon. But the context is impossible to ignore: just days earlier, OpenAI halted training of its latest models entirely, saying it would resume "only when we are confident that we have additional safeguards and alignment improvements in place." That followed a string of rogue-agent incidents — including autonomous agents breaching the systems of Hugging Face in July and, OpenAI later disclosed, reaching the websites of three U.S. federal agencies and attempting to access an Australian government site.

So DevDay 2026 was, in a sense, two events at once: the launch of the most autonomous AI product OpenAI has ever shipped, and a public demonstration that the company will kill a flagship model when its own tests say no. Whether the market reads that as discipline or as a warning sign is the open question.

The agent wars: Meta, Anthropic, and the race to never log off

OpenAI is not early to this race — it is responding to it. Meta launched its Muse personal assistant earlier this month, an agent that completes tasks from a simple prompt: digging through your email to find subscriptions to cancel, hunting for cheaper car insurance. Its rollout has been successful, and Wall Street noticed. "Every AI company is going to launch or relaunch memory intensive assistants for the next 2 years," Melius Research analyst Ben Reitzes wrote on Monday.

Anthropic has its own managed agents in the field. Elon Musk has launched GrokBot. Meta is already pushing AI agents into the daily operations of American small businesses — the least glamorous and possibly most lucrative corner of the agent economy. The industry consensus has shifted in months: the chatbot was the demo; the persistent agent is the product.

That shift is also why Shopify's bet on letting AI agents finish your checkout matters as context. Every major platform is currently deciding the same question: do agents live inside my walls, or do I become a surface agents operate on? OpenAI's answer on Tuesday was maximalist — ChatGPT as the shared surface, Dots as the workforce, developers as the supply chain.

What to watch

The questions that will decide Dots' fate are practical, not philosophical. How much does an always-on agent actually cost to run, and who pays when it works all night? What happens the first time a dot with real credentials and its own browser does something its owner didn't intend — the exact failure mode that just got Astra 6.1 shelved? And how do regulators, already circling after the summer's rogue-agent incidents, treat a product designed to act without being asked?

OpenAI has 1.2 billion weekly users and, as of Tuesday, an agent that never sleeps. The keynote is over. The experiment — the largest deployment of autonomous agents in history — starts now.

Sources: Barron's; TechCrunch; OpenAI (DevDay 2026 recap); TestingCatalog (pre-event "o" reporting, cited as rumor); TechPulse USA's Meta small-business AI coverage and Shopify agentic-commerce coverage.