Sarah Friar, OpenAI’s CFO, published a post on September 8 called “The Work Now Within Reach.” The thesis: as models get more capable and cheaper to run, more work becomes worth doing at all. It’s an investor narrative, weeks out from a public listing. The numbers in it are still worth reading.
The numbers
More than a billion weekly active users. 2.5 million businesses using OpenAI products. The research team clocks 3.1 agent-workdays for every human workday — agents there already put in three times the hours humans do.
On cost: GPT-5.6 Sol cut serving costs by 20 percent, and token-generation efficiency improved by more than 15 percent. Then there’s the in-house chip, Jalapeño, which Friar says delivers 1.5 to 1.9 times more peak token throughput per watt than commercial alternatives, and cuts end-to-end latency by a factor of 1.7 to 3.6.
The loop she builds from it: consumer and enterprise reinforce each other, cheaper AI makes previously impractical work viable, and owning the stack pushes cost and latency down further. Her line: revenue from growing adoption funds further research and infrastructure.
What stands out
3.1 is the number that sticks. Not because it’s large, but because it comes from inside the house and has a definition you can follow. If a research team burns three agent-days per person, “will agents replace people” is the wrong question. The right one is: how many parallel threads can one person actually keep track of? Three works. Thirty becomes a management problem nobody has built a tool for yet.
The rest is balance-sheet poetry with a usable core. A custom chip with roughly double the throughput per watt is what pushes prices down over time — at OpenAI, and through competition, elsewhere too. If you have the Fable 5.1 price list in your head, you know cache reads there just dropped 75 percent. That curve points the same way for everyone.
Sources: OpenAI: The Work Now Within Reach