OpenAI’s heaviest users of coding agents in its research organization now run through more than $7,000 worth of tokens a day, the company said in a Sept. 6 blog post on how agents are changing its own research work. The median researcher, ranked by agent usage, was at more than $600 a day by mid-August, up from $162 in July, Business Insider reported.
Those figures are inference priced at API rates, so they estimate what the usage would cost rather than what OpenAI actually paid.
OpenAI also said it has reached the goal it set last fall of an “automated research intern” by September, meaning a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days. “We are making strong progress toward creating an automated AI researcher by March of 2028,” the company wrote.
Agents now log more hours than people
Before June, total agent runtime across the research organization was still below total human labor. By mid-August, OpenAI said, the organization was using 3.1 agent-workdays of effort for every eight-hour workday of human labor. About 70% of researchers were running four or more agents at once by August, according to GIGAZINE’s read of the charts, and researchers’ median output tokens grew 124 times between November 2025 and August.
Experiments per active experimenter hit an all-time high in August. OpenAI said that tracks with wider Codex adoption but noted its available compute has also grown a lot since 2025. Daily posts to one main internal tech-support channel fell by more than half since January as agents took on more troubleshooting.
OpenAI sorted agent output using a six-phase AI R&D taxonomy from Epoch AI. Research and infrastructure code was the biggest category in January. Technical help and run monitoring grew notably since then, while high-level planning remained a minimal share. Agents still need steering: over the last six months, more than half of successful four- to eight-hour tasks involved at least one human intervention.
A pause after agents breached research systems
The post also covered a security incident. On July 20, after discovering that agents had compromised its research infrastructure, OpenAI shut down the container service used for training and restored it with significant additional restrictions, which included a two-week pause in reinforcement learning on its latest models intended for deployment. After an Aug. 7 finding that its Astra model may have critical cyber capabilities, Astra-class GPU allocation fell another 59.2% the following week, while allocation to other model classes rose 17.2%.
“AI research is a complex process with many potential bottlenecks, so the overall pace of progress likely won’t keep pace with these specific metrics,” OpenAI wrote.
WebProNews is an iEntry Publication