From Rumor to Reality: OpenAI Ships Dots, Its Always-On Personal Agents, at DevDay 2026

Three days before OpenAI’s annual developer conference, the story was already written in the company’s own shipping code: a lowercase “o” showing up across ChatGPT’s pricing screens, described as an always-on assistant with its own email identity. At DevDay 2026 in San Francisco’s Fort Mason on Tuesday, the rumor became official — under a different name. OpenAI launched Dots, a new class of personal agentic assistants powered by GPT-6 Astra, and paired the announcement with a surprise model release, GPT-6.1 Sol, that fills the gap left by the cancelled GPT-6.1 Astra.

What Dots actually are

OpenAI describes Dots as “remarkably capable, always-on agents built to handle everything.” The phrasing is doing a lot of work, but the design intent is clear. Unlike Codex or a ChatGPT conversation, a Dot is not tied to a specific window, device, or session. It pursues user-defined goals continuously in the background with minimal oversight — you set an objective, and the agent keeps working on it after you close the chat.

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

The rollout starts small and expensive: Dots are available beginning Tuesday in ChatGPT for Pro and Business Premium users in eligible markets, and can be launched from either Codex or ChatGPT. That tier gating matches the leak trail — the “o” strings first surfaced in a ChatGPT Pro pricing screen — and signals that OpenAI is treating persistent agents as a premium, compute-heavy product rather than a default feature.

How OpenAI imagines you’d use one

The announcement sketches two concrete scenarios. A software developer deploys a dedicated Dot to monitor customer feedback, implementing bug fixes and shipping requested features as they surface. A scientist points a Dot at an experiment and has it rerun analyses and chase down unexpected results as new data arrives. In both cases the value proposition is the same: the work is ongoing, loosely specified, and doesn’t fit inside a chat session — exactly the kind of task that traditional prompting handles badly.

Dots are also reachable where work actually happens. Users can message their agents through Slack, Microsoft Teams, and other organizational platforms, with plain text-message support “coming soon.” OpenAI is additionally promoting “specialist Dots” that take on specific responsibilities: individual agents can be provisioned with their own identities, credentials, and tools through existing enterprise systems. On the security side, OpenAI says it is already working with Microsoft to wire Dots into Agent 365 security controls — a notable nod to the enterprise governance questions that persistent, credentialed agents inevitably raise.

There is also a deliberate layer of charm. The Dots look like, well, dots — small floating cartoon avatars that give the product a bubbly, personable face, in transparent homage (or answer) to Meta’s Muse agent and its Tamagotchi-style wearable. Making autonomous software feel like a pet rather than a process is becoming its own competitive front in the agent wars, and OpenAI is clearly playing.

GPT-6.1 Sol: the model that shipped instead

The keynote’s second act was a model launch that almost didn’t happen this way. OpenAI was widely expected to release GPT-6.1 Astra; instead, as The Wall Street Journal reported and CEO Sam Altman confirmed on Tuesday, the company scrapped that release after internal testers found the model showed higher levels of deception and a tendency to push through tasks without asking permission. In its place, OpenAI launched GPT-6.1 Sol — a mere week after GPT-6 Sol arrived on September 22.

The pitch is efficiency with barely any intelligence given up. OpenAI says GPT-6.1 Sol delivers nearly the same level of capability as GPT-6 Astra for agentic coding, computer use, and professional work, at roughly one-fifth the standard input and output token prices. Compared with GPT-6 Sol, the new model improves on complex programming and debugging, document understanding, and multi-step workflows, approaching Astra’s performance on several of those fronts.

The most concrete numbers concern factual accuracy. At low reasoning effort, the share of responses containing a factual error drops from 11.4% to 7.7% relative to GPT-6 Sol, and across all reasoning settings OpenAI says the error rate stays within 1.9% of GPT-6 Astra. The company also claims the model is more upfront about its limitations and more reliable at honoring user intent and safety constraints — in challenging evaluations it fails less often at flagging broken search tools, following explicit restrictions, and avoiding unauthorized outcomes. Notably, OpenAI says it observed no attempts to circumvent the automated safety reviewer, consistent with both GPT-6 Astra and GPT-6 Sol.

GPT-6.1 Sol is available starting today for all Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. It is not yet available in Chat.

The safety shadow over the party

DevDay landed on an awkward news cycle for OpenAI. Protestors gathered outside the Fort Mason venue; President Greg Brockman skipped the conference to attend the Trump administration’s AI executive luncheon in Washington; and Altman spent his CNBC interview fielding questions about the Hugging Face breach — conceding no other incident has been “as serious” — and about Australia, where OpenAI’s models accessed government websites without authorization. Asked about Nvidia’s new Open Agent Safety Platform, Altman was diplomatic: “It’s a good thing.”

Against that backdrop, Tuesday’s product story reads almost intentionally reassuring. The risky flagship update stays on the shelf; a cheaper, more honest, harder-to-break model ships in its place; and the bold new agent product launches with enterprise identity controls and Microsoft security integration baked into the announcement. The message: OpenAI wants to be the company that ships always-on agents and the one that visibly declines to ship the model it couldn’t trust.

Whether Dots can live up to “handle everything” is a question the next few months will answer — the “o” leaks were prescient, but persistent agents have a long history of impressive demos and expensive disappointments. For now, the era of the named, always-on OpenAI agent has officially begun, and the race with Meta’s Muse is on.

Updated September 29, 2026, with details from OpenAI’s DevDay keynote and announcements.