OpenAI's research org runs 3.1 agent-workdays per human workday, and the median researcher burns $600 a day

OpenAI published "Research acceleration: the view inside OpenAI" on September 6 with numbers no lab has put in public before (OpenAI). As of mid-August, the research organization uses 3.1 agent-workdays of effort for every workday of human labor. It says it reached its internal goal of an "automated research intern," an agent that carries out multi-day well-defined tasks under human direction, and it's targeting an automated AI researcher by March 2028. The post says agentic systems have contributed to progress toward recursive self-improvement, which is the first time OpenAI has attached a date to that claim.

The spend curve is what I keep rereading. Daily coding-agent spend per researcher sits near $0 in February 2026, about $50 in April, about $150 in June, and roughly $600 by late August at API prices. The 90th percentile exceeds $7,000 a day. Simon Willison read the same chart and pinned the steep late-July inflection to internal access to what shipped as GPT-6 Astra, writing that he's "intrigued at what caused that significant acceleration" and concluding that's the most likely explanation (simonwillison.net). That's an inference, not a disclosure, and Willison frames it as one.

Take the absolute number seriously for a second. $600 a day per engineer in tokens is roughly $150,000 a year on top of salary. A frontier lab has decided that's a normal operating cost. I run agent work daily in my personal projects and I flinch at a $40 day. The gap between those two numbers is the gap between "AI helps me code" and "my job is buying compute and reviewing output."

Now put it next to what OpenAI's chief scientist published the same day. Jakub Pachocki's "An Alien Mind" says no lab has solved alignment and monitoring well enough to keep scaling at maximum speed for much longer, and that he expects and hopes voluntary slowdowns become commonplace until shared safety bars exist (OpenAI). He splits goal alignment from value alignment and admits progress on generalizable alignment may not sufficiently outstrip progress in general intelligence. Altman reposted it with "An important post from Jakub:", pulling about 1.17M views (X).

So one post says the loop is working and we're targeting an automated researcher in eighteen months, and another from the same org on the same day says nobody has solved the thing that would make that safe. TNW found the seam: when safety restrictions forced a 59.2% GPU cut to Astra-class training, other model classes absorbed 85% of the decline, so compute moved rather than disappeared (The Next Web). They also report that over half of successful four-to-eight-hour agent tasks needed at least one human intervention. That last number is the one to budget against. Not 3.1x. Half your long tasks still need you.