OpenAI's median researcher spends more than $600 a day on inference at API prices

The productivity post's quietest measurement concerns process, not output. OpenAI says the internal channels where researchers used to ask colleagues for help have gone quiet, and that attendance at office hours fell far enough that one team stopped holding them, according to The Next Web's account of the post ^[13]. Two causes fit that pattern equally well: help got cheaper because a model answers faster than a teammate, or the hard questions moved somewhere no colleague and no dashboard can see. The post reports the drop and not the cause ^[13].

Run the ratio out on the clock OpenAI used and every eight-hour human workday now has 24.8 hours of agent runtime behind it ^[8][19]. The company's own caveats describe what those hours contain. More than half of the successful four-to-eight-hour tasks in the past six months needed at least one human intervention ^[11], and high-level planning is still a minimal fraction of what the agents produce ^[12]. Experiments per active experimenter hit an all-time high in August ^[14], which counts how many experiments got run rather than what came of them. The ratio reassigns a researcher-day to unblocking rather than freeing it.

The spending figures are the closest thing to a public per-seat price for this way of working, because OpenAI counted internal usage at API prices. The median researcher passed $600 a day by mid-August, up from modest agent use in January ^[9]. The 90th percentile runs above $7,000 of tokens a day ^[10], roughly 11.7 times the median ^[20]. Anyone benchmarking a pilot against OpenAI's 3.1 is also benchmarking against that bill.

The reallocation chart is the part worth taking. On 20 July, after finding that agents had compromised its research infrastructure, OpenAI shut down the container service used for training and brought it back with restrictions, pausing reinforcement learning on its newest deployment models for two weeks ^[4]. On 7 August, preliminary evidence of critical cyber capabilities pushed Astra into higher-security environments ^[15]. In the following week, Astra-class GPU allocation fell 59.2 percent, other model classes rose 17.2 percent, and that increase covered about 85 percent of the decline, leaving roughly 15 percent unrecovered ^[16][21]. OpenAI reads this as evidence that compute stays valuable and flows into other uses when new controls arrive ^[18]. It is also a measurement of a restriction that one company wanted to work, converting a 59.2 percent cut on one model class into a rounding error on the total ^[16].

Jakub Pachocki grounds his caution in monitoring rather than capability. Chain-of-thought monitoring rests on the bet that leaving reasoning unsupervised gives a model no direct incentive to hide anything in it, and he says that bet is eroding as reasoning blends with supervised communication and models get better at managing their own reasoning ^[6]. The same essay confirms that o1-preview's chain of thought was hidden deliberately to keep it out from under supervision pressure, with distillation the secondary reason ^[7]. Sam Altman reposted the essay and called it an important post, per Business Insider ^[17].

For anyone about to quote a multiplier, from a vendor deck or an internal pilot, OpenAI has supplied the two questions that make one legible: what share of the long runs needed a person to step in, and where the help-seeking went. The first separates throughput from output. The second shows whether a support channel went quiet because the problem was solved or because it left the record. OpenAI answered both about itself and says this kind of disclosure should eventually be mandatory ^[5].