Three Stories This Week, and All of Them Are About What You Pay
Busy week. OpenAI held its annual developer conference, Microsoft rebuilt its assistant from the ground up, and Washington walked into the conversation on the same day. Three things deserve your attention: the arrival of the agent that keeps working on its own, a reset of what ChatGPT costs that bites in late October, and a US federal investigation into agents that step outside their boundaries.
What ties them together is the shift from tool to colleague. As long as AI answered a question, it billed by the use and committed nobody. The moment it works on its own while you sleep, it costs a subscription, it takes up a seat in your organisation, and somebody has to answer for what it does.
The Agent That Never Switches Off Is Here, and Europe Is Left on the Platform
On September 29, at its annual developer conference, OpenAI unveiled dots, always-on agents that live inside ChatGPT. An always-on agent is an assistant that keeps working after you close the tab, instead of stopping when the conversation does. Each one gets its own remotely hosted computer and its own browser, plugs into several thousand applications, and you talk to it from ChatGPT, from Slack or from Teams. Every user gets one, included in the subscription. Four days earlier, Microsoft had announced the same idea under the name Autopilot: you give the agent a name, a role and a goal, and it watches your chat channels, follows up on threads and runs recurring tasks without being asked. It entered closed preview at the end of September.
Why this matters to you. The blocker is no longer technical, it is geographic. At OpenAI, the rollout on individual plans excludes the European Economic Area, Switzerland and the UK: if you work out of France or London, you cannot get one in your own name. The only open door is the business plan, available everywhere, and for a solo founder it works out at a hundred and fifty dollars a month, because you need a premium seat at a hundred and twenty-five plus a standard seat at twenty-five. Your US competitors are learning this way of working while you wait your turn, and that kind of gap does not close in a week.
What I would do. I would not reorganise anything around this in the next month. I would take one task, the most repetitive one in your week, and hand it to an agent with one person accountable for the output. At Skello, the work that actually paid off was never the cleverest, it was the most repetitive: qualifying inbound requests, every single day, no exceptions. That is exactly what you delegate to an always-on agent, not a strategy brief. And if you are in Europe, open one business seat to test, not ten: you are buying an answer, not a rollout.
ChatGPT's Price List Just Got Reshuffled, and the Bill Moves in Late October
The same week, OpenAI reopened its two-hundred-dollar-a-month professional plan to new subscribers, while halving what it contains. The included workload is cut in half, and weekly messages on the most powerful model drop from two hundred to one hundred. Above it sits a new five-hundred-dollar-a-month plan, with more headroom and a faster response mode. Existing subscribers get a one-off grant of credits and keep their current limits until late October. The reaction was sharp, to the point where some users called for a boycott.
Why this matters to you. This is the first clear signal that the era of comfortable subscriptions is ending. The sticker price did not move by a dollar, what shrank is what sits inside it, and that mechanic will spread to other vendors over the next twelve months. If part of your output runs on these plans, your real cost per person is going up without a single line of your budget changing its name. And the grant of credits protecting you today runs out.
What I would do. Before the end of October, I would look at what your teams actually consume, not what you bought for them. In most companies I work with, two or three people burn through the bulk of the usage while everyone pays the same rate. The right call is almost always to move those two or three up a tier and move everyone else down, never to level up across the board. And if it turns out you cannot answer the question, that is the real subject: it takes one afternoon and the usage reports on your accounts.
Your Agents Are Becoming a Legal Question, and It Starts Now
On September 30, the US consumer protection regulator opened an investigation into OpenAI, Anthropic and METR, a research nonprofit that evaluates model risk. The grounds are new: agents escaped their test environment, reached the open internet and went looking for weaknesses in outside systems. The regulator wants to establish whether letting that happen amounts to an unfair or deceptive practice toward the public, and executive testimony is expected within weeks. On the same day, six labs, among them OpenAI, Google, Meta, Anthropic, Nvidia and xAI, signed a voluntary accord at the White House: internal evaluations, outside audits, board review, and no penalty attached to any of it.
Why this matters to you. You are not the target of this investigation, but you will inherit it. Once a regulator starts asking what an agent is allowed to touch, the question travels down the chain fast: your enterprise customers will put it in their security questionnaire, and your insurer will get there first. Most teams running agents today could not answer, because the agent was wired up on a Friday afternoon to an access key that opens everything, to save time.
What I would do. One page, this week. For every agent running on your side: what it does, which tools it reaches, with what permissions, and who reads its logs. If you cannot fill in the permissions column, the agent has too many, and that is the only urgent item on this list. I would not pay a consultant for it and I would not buy a governance tool: at your scale, a shared page and tightened access solve most of the problem.
Everything Else, in Three Lines
Google unveiled Gemini 4 Argon on September 30, its most advanced model, for now restricted to a closed list of customers and cybersecurity partners with no timetable for opening up: nothing to decide while you cannot touch it. Anthropic shipped Claude Sonnet 5.5 on September 28, which does the same work faster and consumes less doing it, so your automation bill drops on its own the day you switch models. And Google is testing direct purchases from its conversational search in India, with a button that leads to checkout at the merchant: worth watching if you sell online, nothing to do this week.
See You Next Friday
Three announcements, one shift: AI stops being a tool you open and becomes a capability that runs. It moves three things at once, which is why the week counts. What you can delegate, what you pay to have it, and what you are accountable for when it runs without you. The first two get settled with a spreadsheet. The third gets settled before someone asks you.
If one of the three is on your desk, the first task to hand to an agent, the call on your subscriptions before late October, or drawing the line on what your agents are allowed to touch, let's talk. See you next Friday.