What the 90-hour week at OpenAI does and does not say about your organisation

Will AI make you work more or less? The companies building the technology have been promising a shorter working week for years. Reality tells a different story.

Top investor Marc Andreessen sees exactly the opposite happening in his own circle of friends. Programmers who run multiple coding agents simultaneously and stop going to bed, because sleeping means the work grinds to a halt and the opportunity cost of a night’s sleep has become too high. He calls them AI vampires. We are now seeing the same phenomenon at frontier labs like OpenAI and Anthropic, where a 90-hour working week is not uncommon.

In December 2022, Kamelia Aryafar, then senior engineering director at Google Cloud AI, wrote that by 2025, developers would be able to complete a working week in four days thanks to AI. In April of this year, OpenAI took up the gauntlet with a 13-page policy document, Industrial Policy for the Intelligence Age, calling on governments and employers to trial a 32-hour working week with full retention of salary. The reasoning was that the efficiency gains from AI should benefit the employee, and not exclusively the shareholder.

This week, the BBC spoke with current and former employees of OpenAI, Anthropic, Meta, and Google. At OpenAI, that four-day working week was never trialled, says a former employee. What they did have: crisis meetings, working through the weekends, and performance reviews after which colleagues would disappear overnight. Seventy hours a week was standard. During sprints—the weeks leading up to a major release—working weeks at both OpenAI and Anthropic reach up to ninety hours, according to multiple sources.

We see the same pattern reflected in research. Aruna Ranganathan and Xingqi Maggie Ye from UC Berkeley Haas spent eight months shadowing a US tech company with around 200 employees, who had free access to generative AI without its use being mandated. They conducted over forty interviews. What they observed was published in the Harvard Business Review in February: people worked faster, took on a wider variety of tasks, and stretched their working days, often without anyone asking them to do so. Product managers began to code themselves. Designers picked up data analysis. Engineers, conversely, spent more time reviewing the work of colleagues who could now also code. Breaks dwindled, because there was always time for just one more prompt before lunch.

In short: those who use AI more do not work less, but end up doing much more.

Building and applying are two different things

The easy conclusion is that these companies preach one thing and practise another. The labs are engaged in a capital race where a few months’ lead makes the difference between being a market leader and a footnote. Meta is allocating over a hundred billion dollars to infrastructure this year. In such a race, a 90-hour working week is an understandable, albeit unpleasant, outcome. That tells us something about the conditions under which frontier AI is built. However, it says almost nothing about whether the technology actually saves time in an average business.

Compare it to Ford. The assembly line started running in 1913 and increased productivity spectacularly, but it did not make anyone’s working week shorter. Those gains went into volume and margin. It was not until 1926, thirteen years later, that Ford announced the five-day, 40-hour week with no loss of pay, because management made a conscious decision to do so. Ford knew perfectly well that it benefited him too: well-rested workers who could go shopping on Saturdays were good for business. But it was a decision, not an automatic outcome.

Neil Thompson, an innovation researcher at MIT, puts it dryly in the BBC piece: people assume that 20 per cent less work results in a four-day week, but in reality, new work is created.

What AFAS did differently

And then there is AFAS. Bas van der Veldt announced in 2024 that their Leusden office would close on Fridays from 1 January 2025. Over 700 employees, 32 hours instead of 40, retaining full salary, profit sharing, pensions, and holiday allowances. One hundred per cent pay, eighty per cent time, one hundred per cent productivity.

After a year, revenue and profit have grown by roughly 12 per cent, customer satisfaction has remained the same, employee satisfaction has risen, and absenteeism has fallen. In November 2025, the scheme was made permanent.

What is interesting is the sequence of events. AFAS first removed Friday from the calendar and only then started looking for the financial headroom to pay for it. Meetings without a clear purpose were scrapped. Customer consultations shifted from one-on-one to group sessions, allowing the company to bill in one day what previously took four. AI was integrated into customer service. Corporate anthropologist Jitske Kramer, commissioned by AFAS to research the transition, saw people working with much greater focus as soon as time became scarce. Her diagnosis of the old situation is ruthlessly recognisable: if you have hours to spare, you will just invent another report to write, and that is how we keep each other busy with a load of nonsense.

What you can learn from this

Reinvent the organisation

AI gains are not found in prompting more effectively, but in automating processes that currently require a human behind a computer. Creating these “skills” across all departments frees up time to focus more intently on customers and innovation. For instance, thanks to 25,000 agents in their operations, McKinsey saw a 25% increase in quality time spent with clients.

Seek gains through elimination as well

At AFAS, a significant portion of the extra time came from eliminating meetings and overhauling their service delivery model. Layering AI on top of existing processes primarily just speeds up work that should not have existed in the first place.

Be vigilant against “AI burnout”

Work is no longer glued to the PC; agents can continue running while you are having lunch, and avid Claude and Codex users can now run their coding assistants straight from their phones. This triggers the “AI burnout” effect, where people experience cognitive overload. Scheduling moments of rest for yourself and your colleagues is crucial for sustained, long-term productivity.

Learn to swim by diving in

Technology is no longer the bottleneck; anyone with time and interest can master AI tools. The divide between the AI-literate who optimally deploy these systems and those who occasionally share a prompt with their chatbot of choice is growing by the day. However, you only learn to swim by doing, so use Claude Code, OpenAI Codex, and the equivalents of Copilot or Mistral at least once a day to make something in your own work more effective. Otherwise, you will fall increasingly out of step with both the speed and the work ethic of Silicon Valley.

Final thoughts

At a task level, the promise holds absolutely true. Work that took an afternoon last year now takes twenty minutes. What you do with those remaining hours is entirely up to you.

In 1930, Keynes predicted that his grandchildren would work 15 hours a week by around 2030. When it came to wealth growth, he was surprisingly close. When it came to what we would do with that growth, he was completely wrong. Those who wait for technology to do it for them are waiting for something we have already been waiting for almost a hundred years. Those who make a decision and set clear boundaries will see what happens within a year.