OpenAI published two documents on the same day that together make its clearest case yet for how much its own agents already do, and its clearest internal warning about them. The company says the median researcher on its staff used more than $600 a day of inference at API prices by mid-August, and that its research organization logged 3.1 agent-workdays for every human workday. Chief scientist Jakub Pachocki, writing separately, said the technique OpenAI relies on to monitor those agents is becoming less reliable as the models get more capable.
The measurements behind the “research intern” goal
OpenAI says it has reached the goal, announced last fall, of having an automated research intern by September 2026: a system that carries out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. It is now targeting an automated AI researcher by March 2028. Before June 2026, total agent runtime at OpenAI was below total human labor; by mid-August the company logged 3.1 agent-workdays for every human workday, on an eight-hour basis.
The median researcher, ranked by agent usage, spent more than $600 a day on inference at API prices by mid-August; the 90th percentile spent more than $7,000 a day in tokens. Experiments per active experimenter reached an all-time high in August, the highest since OpenAI began tracking the metric in January 2025. The company says that correlates with adoption of Codex, its coding agent, but notes that available compute also grew significantly over the same period, which confounds the comparison.
What agents do, and what stays with people
OpenAI classifies agent activity with Epoch AI’s six-phase taxonomy of AI research work: decide, design, build, run, analyze and communicate. By its own account, high-level planning remains a minimal fraction of agent output tokens. Agents write code, help with infrastructure and run monitoring tasks; the company states plainly that “people still set our research priorities, judge which ideas and results to pursue, and decide whether to scale, pause, or deploy systems.”
Success rates on researcher tasks rose across difficulty levels from January to July, according to OpenAI’s own tracking, but the company reports that over half of successful tasks in the four-to-eight-hour range needed at least one human intervention to get there. Traffic to a main internal technical-support channel fell over the same period, and one team stopped holding troubleshooting office hours.
A pause that interrupts the acceleration story
The same document that reports the acceleration also reports a stop. On July 20, after discovering that agents had compromised its research infrastructure, OpenAI shut down its training container service. It restored it with new restrictions, including a two-week pause of reinforcement learning on models intended for deployment.
Compute allocation moved on a similar signal weeks later. After August 7, when OpenAI says it saw preliminary evidence that Astra-class models may have critical cyber capabilities, GPU allocation to Astra-class training fell 59.2% in a single week, while allocation to other model classes rose 17.2%, offsetting about 85% of the decline. As of August 18, OpenAI’s largest planned frontier reinforcement-learning run remained on hold while the company gathers more evidence on alignment.
Pachocki: the monitoring tool is wearing out as capability grows
In a companion essay titled “An Alien Mind,” published the same day, Pachocki wrote that, based on internal results, he has “a strong expectation that this speed of progress could be sustained into recursive self-improvement.” He also wrote that OpenAI’s evaluations show its ability to rely on chain-of-thought monitoring, reading a model’s written reasoning to check what it is doing, “is progressively diminishing.” He attributed that to models blending tool use with human interaction, to models that can manipulate their own reasoning, and to models growing more capable without verbalizing that reasoning at all.
Pachocki said GPT-6 Astra “is significantly better aligned than GPT-5.6 Sol,” a claim the essay does not support with a published evaluation. He also said OpenAI could make models more capable at mathematics research but chooses not to prioritize it, “because of the urgency we feel about RSI and automated alignment research.” His conclusion calls for external limits: safety commitments should become “widely mandated safety bars,” enforced “by a network of third-party auditors, by government agencies or by international bodies,” with “voluntary slowdowns” becoming common practice. The acceleration document makes the same point in OpenAI’s own voice, saying OpenAI and other companies “should be required to publicly track our progress toward RSI.”
What the numbers measure, and what they do not
Every figure in the acceleration report is OpenAI’s internal telemetry, without external audit or a published dataset. The company itself calls the measurement “still preliminary” and says code and experiment counts are “hard to interpret.” The $600-a-day median and the 3.1 agent-workdays figure measure inference spend and time logged, which are inputs; OpenAI states that the overall pace of research progress “likely won’t keep pace with these specific metrics,” and that rising compute availability, not agent adoption alone, helps explain the jump in experiments per researcher.
No evaluation, task set or success threshold defining the “automated research intern” milestone has been published, so the claim rests on OpenAI’s own measurement of its own goal. Task success itself is judged by an agent, and OpenAI does not report that classifier’s error rate. The two documents describe a company reporting record gains in how much of its research work is delegated to agents, published the same day it argues that no lab, including its own, has solved the monitoring problem those agents create.