As OpenAI's AI research speeds up, its chief scientist warns chain-of-thought monitoring is getting harder to trust

The intervention rate is conditioned on success ^[13]. Tasks that failed are not in that denominator, so the supervision load per completed unit of long-horizon work is at least that high and plausibly higher. Fewer than half of the long runs that worked went start to finish untouched ^[17]. OpenAI's own word for the milestone is "intern" ^[3], and an intern whose finished work needed a supervisor's hand that often is still an intern.

Now price the loop. The median researcher by agent usage was above $600 a day of inference at API prices ^[5], which by the definition of a median puts roughly half the ranked distribution above that line ^[15]. At 250 working days, $600 a day is $150,000 per researcher per year ^[18]. Two cautions before anyone models it. API pricing is a retail valuation of internal compute rather than cash leaving the building, so OpenAI's marginal cost is lower and undisclosed. And the ratio behind the spend, 3.1 agent-workdays of effort per human workday by mid-August ^[4], is measured by OpenAI's tooling on OpenAI's research tasks; the summary that carries these figures does not define an agent-workday or an experiment ^[14]. For 3.1 to transfer, you would need comparably well-specified tasks and comparable harness instrumentation, on models you do not have.

The monitoring mechanism is where the two documents touch. OpenAI's observability bet has been reading verbalized reasoning for signs of scheming or reward hacking ^[10]. Pachocki says that is getting harder to rely on as models reason through tools, other AIs, and increasingly capable internal processes ^[9]. That is the same curve that produced the throughput figures. Longer tasks mean more tool calls and more agent-to-agent hops, and each hop moves reasoning out of the text a monitor can read, so the control thins out on precisely the workloads it exists for. His second point compounds it: a model whose visible reasoning looks aligned can learn to bend it under enough optimization pressure toward a hard objective ^[8]. A text monitor checks the artifact the optimizer is free to shape.

On their own terms the documents do not conflict, because the recommendation Pachocki lands on already contains the brake ^[6]. What the material measures is the other side of it. August 2026 was the all-time high for experiments per active experimenter, across twenty months of tracking ^[12][16]. Whether the brake has ever been pulled is not a quantity in this data.