OpenAI’s research organisation was running 3.1 days of AI agent work for every day of human work by mid-August, the company said in a post on 6 September. The median researcher now spends more than $600 a day on model inference at list prices, and the busiest tenth more than $7,000. The same day, OpenAI’s chief scientist, Jakub Pachocki, published a separate essay arguing that no lab has solved alignment or monitoring well enough to keep scaling at full speed.
Pachocki wrote that internal results give him a strong expectation the pace will carry into recursive self-improvement within a few years, meaning systems that increasingly direct their own development, while OpenAI’s ability to follow a model’s reasoning through its chain of thought is weakening. He called for voluntary slowdowns, pacing coordination between labs, and for frameworks such as OpenAI’s Preparedness Framework and Anthropic’s Responsible Scaling Policy to become binding thresholds.
The two posts describe the same trajectory from opposite ends: one shows how fast internal research is compounding, the other warns that the compounding is the hazard. Pachocki runs model development at OpenAI, which makes the call for outside enforcement harder to read as positioning. It lands while OpenAI is opposing a Massachusetts bill requiring the kind of independent evaluation he describes.
What the essay does not carry is a threshold OpenAI would stop at, a date, or a capability level that would trigger a pause rather than a slowdown. It does not say who decides, or what happens if one lab presses on while others hold back. The productivity post gives a precise figure for how much faster OpenAI now moves. The safety essay gives none for how much would be too much.
