On 6 September OpenAI wrote that it has an automated research intern and that its research organisation now uses, in total, 3.1 agent-workdays for every human one. The same day its chief scientist Jakub Pachocki published an essay saying no lab has solved alignment well enough to keep going at full speed for much longer.
- Per OpenAI, by mid-August the median researcher was spending more than $600 a day on agents at API prices, and the 90th percentile more than $7,000.
- The next stated goal is an automated AI researcher by March 2028.
- Pachocki wants safe-scaling frameworks to become mandatory bars, enforced by third-party auditors, government agencies or international bodies.
Six hundred dollars a day. That is what the median OpenAI researcher spends on agents, if you count the tokens at the price they sell them to you.
The number comes from a post the company published on Sunday, 6 September. It says the goal announced last autumn has been reached: they have an automated research intern. By intern they mean a system that, under human direction, carries out well-defined research tasks, including ones that would take a skilled person a few days.
The same day, on the same site, their chief scientist published an essay that reads as if it came from a different company.
What they measure and what they don't
The important part is further down the first post: the agents still need a person next to them. In more than half of the successful four-to-eight-hour tasks, someone had to step in at least once. Intern is the right word. It does the work, but you don't leave it alone with the keys to the warehouse.
They also write that such numbers are easy to measure and hard to interpret. More code and more experiments do not automatically mean more discoveries, and their compute has grown alongside the agents. They admit it in the same post, which is honest.
The essay
Pachocki describes how in 2023 he and his colleague Szymon spent the night at the office after the first results with reasoning models. They were not thinking about benchmarks. They were thinking that they would see machines smarter than themselves within their lifetime, and about how to tell people what that means.
Three years later he writes that no lab has solved alignment and monitoring well enough to continue scaling responsibly at maximum speed for much longer. He expects and hopes that voluntary slowdowns become commonplace until shared safety bars exist.
I know what someone will say. Nice words from a man whose company is boasting about its pace the same day. But the essay also contains an uncomfortable admission: by their own evaluations, their ability to rely on reading the models' reasoning, their primary bet, is progressively diminishing.
So we have two sentences from the same house. One measures how fast it moves, the other asks who will pull the handbrake. Which one carries weight will show the first time someone actually slows down without an incident forcing them to.