Dots Runs Always-On Agents on the Astra OpenAI Just Refused to Upgrade

The agent OpenAI would not upgrade is the agent it put on duty. Dots are always-on, named agents powered by GPT-6 Astra — the September 3 model — not by the GPT-6.1 Astra successor canceled the day before for missing the bar on scope, authorization, and honest reporting of its work. Shipping a persistent agent runtime on the older flagship is the product decision. The avatar is not.

DevDay’s other move is distribution. OpenAI’s recap says it is opening ChatGPT as a surface where developers launch native experiences to what it calls a collective 1.2 billion weekly users. That usage figure is the company’s. It is not an audited reach number.

What Dots actually are

OpenAI’s launch post (September 29) describes Dots as agents with their own cloud computer and browser, able to keep several projects moving and to connect, through plugins, to over 4,000 apps. You can open the dot’s computer and watch. A dot may also be allowed onto your laptop. Reach is ChatGPT on desktop, web, and mobile, plus Slack and Teams, with texting described as coming soon. The first dot is included at no extra charge on eligible Pro and Business Premium plans. Conversations with the dot do not count toward ChatGPT usage limits; tasks it starts in Codex or ChatGPT Work do. Enterprise, Edu, and Healthcare can try a beta only if a workspace admin turns it on. It is off by default.

The permission design is the part worth reading against the Astra hold:

• Proactive research while you are away uses connected apps with read-only tools. It cannot send messages, change app content, or drive the browser or computer until you approve a write.
• Custom Rules allow, require approval, or block specific actions. Built-in safety rules still apply.
• Auto-review checks actions that could touch accounts or share information against your instructions, those rules, and safety requirements. Changing a password stays with the user.
• Saved passwords for supported sites are not shown to the model.
• OpenAI says monitoring can pause or stop a dot, and that dots can still make mistakes on consequential work.

Specialist dots — a separate identity, credentials, and a defined job — are an enterprise pilot, not the consumer default. OpenAI says it is working with Microsoft so they can sit under Agent 365. Internal tests it names include procurement, invoices, and support. No outside case study with an error rate shipped with the post. NBC News quotes Sam Altman calling a dot “an AI helper that always has your back,” and calling Dots a premium product because of compute cost.

ChatGPT as the store

Folded into the same DevDay, not a second product story:

• Plugin extensions put a developer’s app in the sidebar, in panels beside the chat, or as a file viewer. Users still choose plugins and approve access. The Next Web names Canva, Figma, and Adobe as early examples.
• Sign in with ChatGPT lets people take plan usage to 16 partners, including Cognition’s Devin, Notion, Vercel, T3, OpenClaw, and Dactyl, and cap what each app can spend. OpenAI says identity works globally; plan usage is for Plus and Pro in participating tools. The Next Web adds that those apps do not receive the user’s ChatGPT conversations or memories.
• OpenAI Marketplace: eligible enterprise customers can apply part of an existing OpenAI commitment to approved partner software. The recap’s first 32 partners include Figma, Adobe, Sierra, Decagon, HubSpot, Salesforce, ServiceNow, Harvey, Legora, Palo Alto Networks, CrowdStrike, and Baseten. The Next Web reports the beta is for eligible US enterprise customers.

The Next Web, citing the company, also reports over 35 million people building with Work and Codex weekly. That figure does not appear in the recap text. Treat both usage claims as company-reported.

Where it does not launch

OpenAI’s own posts say only “eligible markets.” The press accounts do not match.

The Next Web says an OpenAI statement to press kept Pro users in the European Economic Area, Switzerland, and the UK out of the initial rollout. NBC News quotes Altman more broadly: Dots are unavailable in Europe and the United Kingdom because of the region’s rules. “We have a great track record of getting our stuff launched in the E.U., but they do have a process that takes longer, and sometimes unpredictably longer,” he told NBC. Switzerland is in The Next Web’s Pro exclusion and absent from that NBC sentence. Whether Business Premium is inside the carve-out is not reconciled in the two stories.

The carve-out matters because the product’s risk is not chat. It is an agent with a computer, credentials, and a background loop. That is the setting EU and UK rules are written for, and the setting OpenAI just said 6.1 Astra could not be trusted to stay inside authorization.

NBC also reports that OpenAI’s note the evening before, on the June Australia incident, said “We are sorry and working to do better in the future,” and that staff described those runs as internal models not meant to ship. Dots is a commercial product on the Astra that did ship. This is a different layer than the Agents API, which productized the Codex harness as a billable developer session. The wiki breakout is why the read-only default and auto-review are the story, not the illustration style.

Limits

• No independent measurement of how often auto-review catches a bad action, or how often proactive research stays read-only in practice.
• 1.2 billion weekly users and 35 million Work/Codex users are company figures. The second is via The Next Web, not the recap body.
• Europe availability conflicts: Pro-only exclusion including Switzerland (The Next Web, citing an OpenAI statement) versus a broader Europe-and-UK hold (Altman to NBC). We did not re-read a Help Center matrix.
• Specialist dots and Agent 365 are pilots and an integration goal, not a generally available enterprise control plane.
• Altman’s stage quotes are NBC’s reporting of the keynote, not a transcript we hold.

Sources

• OpenAI: Introducing dots (September 29, 2026)
• OpenAI: DevDay 2026 recap (September 29, 2026)
• TechCrunch: OpenAI launches Dots (September 29, 2026)
• The Next Web: DevDay plans, Pro exclusion, usage figures (September 29, 2026)
• NBC News: Dots launch amid safety questions (September 29, 2026)
• Times of AI: Agents API public beta (September 10, 2026)
• Times of AI: agents and the German wiki breakout (September 4, 2026)