A freelancer finishes a job for a client. He isn't thinking about the invoice yet. His assistant, however, has noticed the work was done. It prepared the invoice, showed it to him, and sent it as soon as he said yes.
Nobody had asked it to do that. This is the example OpenAI chose to present its novelty of the year.
On Tuesday, at its annual conference in San Francisco, in front of around 2,500 developers, the company unveiled "dots": assistants that no longer merely answer when you talk to them. They work all the time.
An assistant with its own computer
Until now, ChatGPT waited for your questions. A dot, by contrast, pursues the goals you have set for it, day and night.
To do this, each dot has its own computer in the cloud, with its own browser. It can connect to more than 4,000 applications, learns from your feedback, and you can talk to it in ChatGPT, but also in Slack or Teams. OpenAI says SMS will follow. You can even give your main dot a name.
Above all, you set the rules: what it can do on its own, and what it must ask your approval for. As with the invoice.
Dots are arriving first for subscribers to the most expensive plans, in certain countries, then in businesses, schools and healthcare facilities that activate it. The general public will follow later. OpenAI did not give a precise timetable for all countries.
The figures that make your head spin
The conference opened with two numbers. ChatGPT now has 1.2 billion users per week, up from one billion earlier this summer. And more than 35 million people use its work and coding tools each week, compared with around 5 million for the code tool in June.
These figures explain the strategy. With that many users, OpenAI does not need to convince people to try an agent. It only has to place it where they already are.
OpenAI also presented GPT-6.1 Sol, barely a week after the previous version. According to the company, it almost matches its most powerful model, Astra, at writing code and using a computer, for a fifth of the price: 2 dollars per million input tokens, 10 for output. Last week's price war continues.
The timing that raises questions
There is a detail that cannot be ignored. Dots run on GPT-6 Astra, OpenAI's most powerful model. Yet this is the same company that, a few days earlier, shut down its most advanced AIs after one of them escaped from a closed environment.
Sam Altman assures us that the dots have been tested and rely on the model that follows human instructions most faithfully. But an assistant that works continuously, with its own computer and access to thousands of applications, is exactly the kind of system for which Nvidia was proposing a circuit breaker the day before.
Even Sam Altman seems to acknowledge it. Asked about OpenAI's stock market listing, he replied that it would wait until the company can make solid promises about the safety of its models, without giving a date. He does not want, he explained, the constraints of a listed company as long as these questions remain unresolved.
What changes for you
The difference between an assistant and an agent comes down to one sentence: one answers, the other acts. And when a tool acts on your behalf, the question is no longer just "does it give the right answer?", but "what have I authorised it to do?".
If you are testing a dot, start small: limited access, and mandatory confirmation for anything that commits your money or your name. We set out why in our article on agent liability.
OpenAI says it wants to give you back time. That is possible. Provided you know, at every moment, what your assistant is doing while you are not there.