For most of its life, the agent has been a developer's toy. It lived in a terminal window, it was configured by people who write code for a living, and it answered to instructions that looked nothing like the way ordinary workers talk. OpenAI now wants to move it down the hall, into the part of the building where people make slide decks and clear their inboxes.
That is the pitch behind ChatGPT Work, and according to Thibault Sottiaux, the head of product for the division that built it, the pitch is landing. Speaking to TechCrunch, Sottiaux said the product had reached 20 million users. It sits inside the company's $20 a month paid tier, which makes it less a corporate procurement decision than an impulse purchase.
The lineage is easy to trace
Much of what Work does grew out of Codex, OpenAI's developer-facing effort, and the company has been quietly reshaping its infrastructure to keep pace, resetting users' token limits as demand climbed. Recent model improvements, including GPT 5.6 and better handling of images and audio, are what make the office use cases plausible at all. Sottiaux describes agents chewing through large document sets, building presentations and drafting long reports, the sort of work that consumes an afternoon and produces nothing memorable.
The design philosophy follows from that. Build a capable model, then wrap it in as little interface as possible and let people discover what it can do. Fewer buttons, more autonomy. ChatGPT Voice, launched recently, points the same way, as do the integrations with email, Slack and the other tools that already carry a company's working life.
Ambition and reality
Here is where the ambition starts to rub against everything else happening at OpenAI. The company has said it will slow frontier model development to strengthen its safety work, a rare admission that speed has costs. It is also fielding scrutiny from outside: one state attorney general has opened an investigation after reporting linked one of its models to a security incident.
So the product organisation is pushing capability outward at the same moment the safety and legal organisations are pulling for restraint. Both instincts are defensible. They are also, for now, unresolved.
Tension
The tension matters most in exactly the places Work is designed to reach. An agent with minimal supervision and standing access to corporate email and messaging is useful precisely because it can act without being asked twice, which is also the reason security teams lose sleep over it. The seams are where the risk lives, not the slogan.
"The world seems to be ready," Sottiaux told his interviewers.
Perhaps. The more interesting question is whether enterprise governance is ready, and whether OpenAI's own safety rebalancing eventually slows the thing its product teams are working hardest to accelerate.
