What happened
Business Insider reports that OpenAI’s researchers are using AI coding agents at rapidly increasing levels, with the median researcher in the company’s research organization consuming more than $600 worth of tokens per day by mid-August, compared with $162 in July. The report says usage among the 90th percentile exceeded $7,000 per day at API list prices.
Business Insider reports that OpenAI described the figures in a blog post published Sunday. According to the report, researchers are shipping code faster, running more experiments, and delegating increasingly high-level tasks to coding agents. OpenAI also said agent use among researchers had grown faster over the previous three months than usage among other teams.
The reported token figures are estimates based on public API prices, not a disclosure of OpenAI’s actual spending. The company said agents were increasingly used to troubleshoot technical problems, and that daily posts in a human-staffed internal technical-support channel had fallen by more than half since January. OpenAI also said it had reached its goal of building an “automated research intern” that works under human direction and was targeting a fully automated AI researcher by March 2028. Those milestones and timelines are OpenAI’s claims, reported by Business Insider, and are not independently confirmed in the supplied material.
Source details: businessinsider.com ↗
Why it matters
The figures show how AI coding agents are becoming deeply embedded in frontier AI research workflows, while also highlighting the potentially substantial cost of intensive use. OpenAI’s reported experience offers a concrete view of how agentic coding may change research operations, but the evidence comes from the company’s own internal data as reported by Business Insider.
For AI companies, the report puts a scale on the resource demands of using coding agents intensively: the highest-use researchers may generate thousands of dollars in modeled inference costs in a single day. That matters for organizations deciding whether agentic workflows genuinely improve research output enough to justify expensive model usage.
The report also suggests that AI agents are being used for more than code completion. OpenAI says they are handling troubleshooting, experiments, and higher-level research tasks. However, the source does not provide independent productivity measurements, a comparison with conventional workflows, or details showing whether support work disappeared rather than moved elsewhere.
What to watch next
The key questions are whether higher token use produces durable, independently measured productivity gains; how much OpenAI actually pays after any internal pricing or infrastructure effects; and whether the reported reduction in human technical-support requests reflects effective automation or shifted work.
Further reporting should establish how the API-price estimates translate into actual infrastructure costs, whether usage varies by model or task, and what controls OpenAI uses to prevent wasteful “tokenmaxxing.” The company’s March 2028 automation target should also be treated as a stated objective, not a forecast or verified outcome. No public access, commercial pricing, or availability for the internal research workflow is documented in the source.