AI Was Sold as a Shortcut to a Four-Day Week. Its Builders Are Working 90 Hours.

OpenAI and Anthropic want every office worker to clock in three or four days a week, powered by AI. The people building the AI still clock fifty, sixty, ninety.

OpenAI published a policy blueprint that included shorter workweeks as part of its plan for managing "AI upheaval." Anthropic's leadership has argued the same case in congressional testimony and op-eds. Cerebras CEO Andrew Feldman, whose company sells the chips powering many of those AI systems, summed up the position bluntly: it is "mind-boggling" that anyone thinks you can work 38 hours a week and still be successful.

Inside those companies, the picture is the opposite. Former OpenAI and Anthropic employees told the BBC that the standard workweek at OpenAI runs 50 to 60 hours, with code sprints and weekend meetings layered on top. Some described weeks pushing 90 hours, particularly around model launches. OpenAI president Greg Brockman has publicly documented his own 90-hour weeks in past posts, a personal style the company has not discouraged.

The pattern shows up in the AI-written code claim, too: an engineer told Fortune that "upwards of 100%" of his normal work is now done by AI tools. The hours he is putting in are roughly the same. The scope of what he can ship has simply expanded.

It is not just an AI-lab problem. It is a pattern with a name, and it has been measured. In February 2026, researchers at Harvard Business School published a study finding that workers using AI tools tend to take on a broader scope of work, not a shorter day. The headline stat in most coverage, that some employees at the AI Work Institute report saving as much as 11 hours per week, sits inside the same finding. The 11 hours are real. What fills them is more work, not more rest. AI compresses the time it takes to finish a task; it does not shrink the workload assigned to the person who finishes faster.

The mechanism is what researchers are now calling "productivity theater": workers get faster, so managers give them more, and the visible metric, output shipped, climbs even as the human on the other end works the same hours or longer. The companies building AI tools are the most extreme case, because they are competing to ship each new model before a rival does. The same dynamic, the HBS data suggests, is already showing up across knowledge work more broadly.

There is a second reason the contradiction is not hypocrisy. AI labs have a financial incentive to push the four-day-workweek story even when their own employees do not live it. A future where AI compresses every office job into four days is also a future where every office customer needs more AI to stay competitive. The shorter the promised workweek, the more compelling the pitch. The longer the actual workweek at the company making the pitch, the more aggressive the model release cycle has to be to justify it.

This leaves a reader in an awkward position. AI tools are genuinely useful. They cut the time it takes to draft, code, and analyze, and the HBS research is clear that the speed gains are real. The question is not whether the work is faster. It is what gets done with the saved time. The same week an OpenAI policy paper argued for a four-day future, the company was shipping its next model on a code sprint that consumed its employees' weekends. The two facts are not in conflict. They are the same strategy viewed from two sides.

The question to ask of any AI rollout is not whether the tool saves time. It is whether the saved time comes with a written commitment about what fills it. A real four-day pilot has to specify three things. First, the workload: a hard cap on what is on the plate at the start of the week, not a vague promise to "do the same with less time." Second, the audit: a way to see whether the saved hours go to other work, with the worker's calendar as the source of truth. Third, the metric: a definition of success that does not quietly default to "more output per person." Without those, a four-day week is just a five-day week with one day of weekend work moved onto the calendar.

The pitch from OpenAI, Anthropic, and Cerebras is not a lie. The hours at those companies are not a secret. Both can be true at the same time, and that is the part the marketing copy tends to leave out. AI is making knowledge workers faster. The companies telling you so are betting that the speed shows up in their own output first, and in your hours last. The order in which those two things happen is the only thing that distinguishes a real shorter workweek from productivity theater with a launch event.