The number OpenAI chose to publish
OpenAI said on 6 September that it has reached the goal it set itself in autumn 2025: an “automated research intern”, a system that carries out well-defined research tasks under human direction. In a post describing how its own research organisation now works, the company said that as of mid-August it consumes 3.1 agent-workdays of effort for every workday of human labour, measured against a standard eight-hour day.
Before June 2026, the company said, total agent runtime across the research organisation was still below total human labour. The crossover took roughly two months.
What the ratio does not mean
OpenAI attached its own caveat, which is worth repeating because the headline number invites the wrong reading. The ratio does not translate into a threefold productivity gain: agent work runs in parallel, duplicates itself, and often needs heavy human steering. The company also said the system does not set research agendas, decide which unanswered questions deserve resources, or judge the significance of a result. Those remain human jobs.
What it costs
The spending figures are the part that generalises beyond OpenAI. By mid-August the median researcher was spending more than $600 a day on inference through coding agents, and the 90th percentile more than $7,000 a day. August set an all-time high for experiments per active researcher since the company began tracking the measure in January 2025.
Purchased computation has become a line item in the daily work of a research scientist rather than a capital decision made once a year. As agent-assisted coding gets cheaper, OpenAI said, the binding constraints move elsewhere — to experimental compute, data quality and research judgement.
Safety incidents show up in the compute logs
The post is unusually specific about how safety decisions bite. After agent infrastructure compromises on 20 July, reinforcement learning training compute fell sharply. When OpenAI imposed restrictions on Astra on 7 August over its cyber capabilities, allocation of Astra-class GPUs dropped 59.2% while allocation to other models rose 17.2% — the work moved rather than stopped.
The company also repeated a limit on its own ambitions: “We do not yet know how to safely get all the way to aligned, full RSI,” it said, referring to recursive self-improvement, and said it would slow or stop where risks became unacceptable.
What to watch
The next stated milestone is a full automated AI researcher by March 2028, roughly 18 months out. The same week, OpenAI’s chief scientist Jakub Pachocki published an essay arguing that no lab has solved alignment well enough to keep scaling at maximum speed. Both documents came from the same company on the same day, and the tension between them is the story to follow.
