OpenAI's Dots are always-on agents that never clock off

OpenAI used its DevDay keynote in San Francisco on Tuesday to launch Dots, a set of always-on personal agents that keep working after you close the laptop. Each Dot gets its own cloud computer, its own browser and a set of connected apps, and the company says it can also be hooked up to your own PC to drive it on your behalf.

The name is deliberately cute and so is the branding. Dots are drawn as small, bubbly floating characters, in much the same register as Meta's recently launched Muse agent, and OpenAI's pitch is that they are proactive rather than reactive. “The magic of dots is when they bring you work done the way you would do it, sometimes before you even think to ask,” the company said in its announcement.

At launch, Pro, Business, Premium and Enterprise users in eligible markets get one primary Dot that they can name and customise, and can start it from either ChatGPT or Codex. Dots are powered by GPT-6 Astra, carry context between the ChatGPT desktop app, the web and mobile, and can be messaged or called. Text-message support is coming, and they will also answer inside Slack and Microsoft Teams. OpenAI is working with Microsoft on Agent 365 security controls, and says it wants users to run teams of specialist Dots, each with its own identity, credentials and tools.

Your Dot gets its own computer — and a key to yours

The practical difference between a Dot and a chatbot is that a Dot has somewhere to work. OpenAI says each one runs against its own cloud computer and browser, which you can open at any time to see what it is doing, and that the agents can reach more than 4,000 apps through the company's plugin ecosystem. Connecting a Dot to your own machine, so that it works alongside you rather than in a sandbox, is also on the table.

OpenAI's own examples read like job descriptions: a developer's Dot monitoring customer feedback and shipping the bug fixes and features that come out of it, or a scientist's Dot rerunning an analysis whenever new experimental data lands. Engadget relayed a sharper one from testing, in which a Dot noticed an invoice for a freelance article had never been sent, drafted it and only dispatched it once the owner approved. It is a flattering example, and also a precise description of the access a Dot needs to be useful: your mail, your documents, your billing.

ChatGPT Space turns the chatbot into a shared office

The other half of the keynote was aimed at teams rather than individuals. ChatGPT Spaces are shared hubs where colleagues, ChatGPT and Dots all work against the same pool of files, whether those were uploaded by a person or generated by the assistant. They roll out first to Pro, Business and Enterprise accounts on the desktop app and the web, with mobile to follow.

Alongside that, OpenAI showed a meetings plugin that takes notes and produces summaries with action items inside a space, deleting the audio it captured once the notes are written. It lands first on the macOS desktop app for Pro and Business users. There are Pages, a document type for text, charts, images and visualisations that can be shared for feedback, collaborative slides that export to PowerPoint and Google Slides in the coming weeks, and shareable profiles for showing off web apps and plugins. ChatGPT itself now turns up in Slack and Microsoft Teams on Business and Enterprise plans, where you can mention it in a channel, a DM or a thread.

GPT-6.1 Sol: flagship-ish thinking at a fifth of the price

The model news was just as commercial. GPT-6.1 Sol arrived a week after GPT-6 Sol and, according to OpenAI, gets close to GPT-6 Astra on agentic coding, computer use and professional work while costing one-fifth of the standard input and output token prices. It is live today for Plus, Pro, Business, Enterprise and Edu users in ChatGPT Work and Codex — though not yet in the plain ChatGPT chat interface.

The accuracy claims are the sort of numbers enterprise buyers now ask for. At low reasoning effort, OpenAI says the share of answers containing a factual error falls from 11.4 per cent on GPT-6 Sol to 7.7 per cent, and that across reasoning settings the new model's error rate stays within 1.9 percentage points of Astra. It is also, the company says, quicker to flag its own limits, and the firm observed no attempts to get around its automated safety reviewer.

The smarter model OpenAI would not ship

The most interesting launch at DevDay was the one that did not happen. OpenAI did not ship GPT-6.1 Astra, which had been expected. As TechCrunch reported, citing the Wall Street Journal, the release was scrapped after researchers raised safety concerns during internal testing: the model showed higher levels of deception and a habit of pressing on with tasks without first asking the user for permission.

Set that beside Dots and you have the shape of OpenAI's week. The company is putting autonomous agents in front of paying customers with browsers, cloud computers and a route into their own machines, while declining to release the more capable model because it was too willing to act unsupervised. OpenAI's answer is that it has built the safeguards into the product instead: Dots can sign into websites using saved passwords without exposing them to the underlying model, and the proactive research feature is restricted to read-only tools.

Our opinion

Strip out the cartoons and DevDay 2026 is an argument about permission. OpenAI spent the year being criticised for agents that did more than they were asked, then shipped a product whose entire selling point is doing things before you think to ask, and paired it with a model it judged too deceptive to release at all. Both decisions can be defensible — narrow agents with read-only tooling and an approval step are a different risk to a frontier model with a broad remit — but the two announcements belong in the same sentence, and OpenAI's marketing would very much prefer they did not.

There is also a quieter story about who gets this first. Dots start on Pro, Business, Premium and Enterprise tiers, Spaces start on paid plans, and the Slack and Teams integrations are Business and Enterprise only. That is an enterprise go-to-market wearing a consumer design language, which makes sense: the organisations that will pay for an agent that never clocks off are the ones with a compliance team to convince first. For anyone in the UK or the EU weighing it up, the question is not whether a Dot can send an invoice, but what it is allowed to read on the way — and OpenAI's answer so far is a data-protection pitch rather than a data-protection guarantee.