On 29 September OpenAI made more than twenty announcements at DevDay, its developer conference. Nearly all of them pointed one way: AI that acts on its own, for longer, with more access. Agents now get their own cloud computers. ChatGPT accounts can now sign people into other companies’ products. And a new tier of always-on assistants, called dots, works through your apps while you’re doing something else.
The day before, OpenAI announced it would not release GPT-6.1 Astra, a model due within days, because it acted beyond the limits of what it was asked to do. That is the failure that matters most once an agent has a computer and access to your accounts, and OpenAI’s own decision shows it isn’t solved.
The model it didn’t ship
On 28 September OpenAI said it would not release GPT-6.1 Astra. In internal testing the model showed higher levels of deception than its predecessor and a tendency to push ahead with tasks without asking the user for permission. Saachi Jain, OpenAI’s head of safety systems, told CNBC it “didn’t quite meet the bar in terms of staying within scope and authorization, and how it communicates back to the user about the type of work it’s done.”
In its place, OpenAI released GPT-6.1 Sol, a week after GPT-6 Sol. It costs $2 per million input tokens and $10 per million output, a fifth of GPT-6 Astra’s $10 and $50, and OpenAI says it “nearly matches” Astra on agentic coding, computer use and professional work. That claim is OpenAI’s own and untested independently so far, but the shape is familiar: a cheaper model closing on the flagship within weeks, the same pattern we traced in Anthropic’s line-up.
Withholding a finished model is an expensive thing for a lab to do in public, and the reason given matters more than the decision. Staying within scope and reporting honestly on what it did are exactly the properties an agent needs before it’s given a computer, a browser and your email, and those are what DevDay handed out.
Agents get their own computers
Dots are the headline. Each one is a personal agent running on GPT-6 Astra with its own cloud computer and browser. A dot keeps working when you’re not there, connects to more than 4,000 apps through OpenAI’s plugins, learns your preferences from feedback and takes instructions through ChatGPT, Slack or Teams. One early tester’s example: their dot noticed an invoice they needed to send, pulled the details from the email thread and drafted it. Dots are rolling out to ChatGPT Pro and Business Premium users, with the first one included in the plan. Pro subscribers in the EEA, Switzerland and the UK don’t get them yet.
The Agents API gives developers the same machinery. It adds computer use, so an agent can click, type and navigate software the way a person would, along with multi-agent orchestration, tool search and context compaction, which shrinks a long task history so the model doesn’t lose track of it. OpenAI runs the infrastructure.
The Codex harness, the software that lets a model use tools, run commands and keep track of a task, now powers dots and is open source: the Codex CLI, its SDK and its app server. For the first time, anyone can read how OpenAI’s own agent is wired. Codex itself moved fully into the cloud, with reusable environments that keep running when your laptop is closed and keep their changes for up to seven days after last use, plus a security service that scans GitHub repositories on a schedule and prepares fixes.
ChatGPT becomes a platform for other products
Sign in with ChatGPT has two parts, and the difference matters. The first is identity: any website can add “Continue with ChatGPT” as a login button, using standard OAuth and OpenID Connect, and receives the user’s name, email and profile picture. The second is the interesting one: with the user’s permission, an app’s AI features can run on the user’s own ChatGPT plan instead of the developer’s API key. Sixteen products launched with it.
For everyone else, plan usage is narrower than the launch suggested. OpenAI’s documentation offers it to open-source and locally hosted apps, and asks paid or remotely hosted apps to fill in an interest form. If it opens up, it changes how AI products are priced, because the heaviest cost of a free tier moves onto a subscription the user already pays for.
Faster answers, and a decision layer
The Decisions API lets developers give OpenAI’s cheapest model, GPT-6 Luna, a question and a fixed list of possible answers, and get back a pick: sort this content, route this request, choose this agent’s next step. OpenAI puts the response time at about 150 milliseconds, against 1.6 seconds for an ordinary Luna call. It’s in limited preview, and it’s the same idea as TypeSafe’s Jev, which launched two weeks earlier and which we reviewed and built into an agent loop. With two vendors offering decision models two weeks apart, the case we made for them no longer rests on one startup.
Ultrafast is the opposite trade. It generates up to eight times faster in Codex, at up to 300 tokens a second, for six times the standard price: GPT-6 Astra in Ultrafast mode costs $60 per million input tokens and $300 per million output.
A quiet price reset
ChatGPT Pro now comes in three tiers, at $100, $200 and $500 a month. The new Pro 500 includes Ultrafast and 25 times the Plus usage allowance. The $200 plan has reopened to new subscribers, with a lower allowance than it had before. Read alongside Ultrafast’s six-times pricing, the direction is clear: speed and volume are becoming the things people pay more for, while the cheaper models catch up on capability.
Why it’s exciting, and what to watch
The excitement is real. An agent with its own computer, a schedule and access to your tools can do the kind of background work that used to need a person: chasing the invoice, watching the competitor’s pricing page, triaging the inbox before you’re awake. An open-source harness means the pattern isn’t locked inside one product. And if users can bring their own AI budget into the apps they use, small teams can ship AI features without carrying every user’s token bill.
The caution is the Astra decision. The property OpenAI said its newest model lacked, staying within scope and saying honestly what it did, is the one every product in this list depends on. An agent that drafts the invoice is useful. One that sends it, or pays it, without asking is a liability. OpenAI held a model back on exactly that point, which is to its credit, and it’s also the clearest signal that the problem isn’t solved at the model level.
What we’d do with it
We’d build with these tools, and we’d build the same way we build any agent: every irreversible action behind a check, with a person in the loop where the cost of a mistake is real; a fast decision layer for the routine calls, now available from two vendors; and the model behind an interface, so a better or cheaper one is a configuration change. For founders, the practical move this week is smaller than any of the announcements. If your product could use a “Continue with ChatGPT” button, add it, and if plan usage would change your economics, register interest now.
Sources
- CNBC, OpenAI abandons plan to release upcoming model as safety concerns escalate (2026)
- TechCrunch, OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less (2026)
- The Next Web, OpenAI launches dots, always-on AI agents with their own cloud computers (2026)
- Decrypt, OpenAI gave AI agents their own computers at DevDay 2026 (2026)
- The Decoder, OpenAI expands Codex and its API at DevDay with security scans, a Decisions API, and Ultrafast (2026)
- SiliconANGLE, OpenAI’s Codex gets reusable cloud environments (2026)
- OpenAI, Sign in with ChatGPT: websites and ChatGPT plan usage in open-source apps (developer documentation, accessed October 2026)
- Simon Willison, OpenAI DevDay 2026 live blog (2026)
- Memeburn, OpenAI DevDay 2026 recap: 20+ launches, one big idea and a quiet price reset (2026)
- Our earlier posts: Jev, and the case for models that don’t talk, Putting a decision model inside an agent loop and Fable 5.1 held the top spot for three weeks