OpenAI's ChatGPT Work Targets White-Collar Workers and Already Counts 20 Million Users

OpenAI pushes agents beyond developers

OpenAI has launched ChatGPT Work, a platform that extends the kind of autonomous AI agents previously tied to its Codex coding tool to ordinary office workers. In an interview with TechCrunch published on August 25, Thibault Sottiaux explained the thinking behind the release. Sottiaux leads OpenAI's core products — the API, agent infrastructure, enterprise offerings, all of ChatGPT including ChatGPT Work, and Codex — and reports to Greg Brockman. As TechCrunch notes, Codex users may already know him as the person who resets their token limits whenever the engineering tool hits a growth milestone.

Sottiaux said the goal was to take agent technology that had been tested on developers, an audience willing to put up with rough edges, and repackage it once it was mature enough for everyone else. ChatGPT Work ships as part of the $20-per-month Plus plan and is available on mobile and web, which he described as part of making the technology reachable by as many people as possible.

Twenty million users already

According to Sottiaux, ChatGPT Work has just crossed 20 million users, which he offered as evidence that the market is prepared for agents that handle complicated tasks with little supervision. The TechCrunch interviewer pressed him on the point, citing Wharton professor Ethan Mollick's observation that ChatGPT Work aims to feel like magic, while Anthropic's Claude Cowork instead surfaces explicit choices — such as A/B tests — and asks the user to decide. Sottiaux pointed to the adoption numbers rather than contesting the comparison.

Discovery as a product philosophy

Asked how you design a product meant to do everything, Sottiaux described the process as partly exploratory: as model capabilities advance, OpenAI learns what the models are genuinely good at and then builds around those strengths. He cited GPT 5.6 as a step change in general office work, including processing large volumes of documents, producing polished slide decks and reports, and carrying out in-depth research.

On design, he argued the interface should stay minimal so the model's capabilities come through directly, rather than hiding them behind buttons and menus. ChatGPT Voice, which he said has grown quickly, is part of that trajectory: the product should adapt to people rather than forcing people to learn the product.

The economics of utility

TechCrunch noted that analysts see an economic motive in the launch — OpenAI wanting to own the application-layer relationship with users. Sottiaux reframed the question in terms of utility, arguing that the more value ChatGPT delivers, the more naturally users are willing to pay for it.

On the gap between what subscribers pay and the compute they consume, he pointed to efficiency work. OpenAI recently announced 80% price cuts under the name Luna, which he characterized as permanent rather than promotional, and said the aim is for the same dollar to buy more capability over time — the work you can do today should cost less to redo six months from now.

Email, messages and trust

The interview also covered a sensitive rollout: access to iMessages, which the interviewer said had shipped that same day, adding to existing email access. For users uneasy about granting an agent a window into private communications, Sottiaux pointed to OpenAI's investment in safety and its practice of publishing benchmark results on model behavior.

Why it matters

ChatGPT Work is OpenAI's bid to turn agentic AI from a developer novelty into a default tool for white-collar work, the segment where AI's effect on jobs is most contested. A reported 20 million users suggests the diffusion is happening fast, and at $20 a month with what Sottiaux calls lasting price cuts, the barrier to adoption is low enough to normalize quickly. The contrast with Anthropic's approach — OpenAI betting on seamlessness, Anthropic on explicit user control — frames a real design debate about how much autonomy workers will accept. And the move into email and messages means trust, safety verification and data handling, not raw capability, are now the factors most likely to determine whether this sticks.