OpenAI's coding agents are now doing 3x the work of human researchers

What happens when your research team effectively triples its headcount overnight, but each new digital hire burns through hundreds of dollars in compute every afternoon?

That is the reality inside OpenAI right now. On September 6, the company published an inside look at how coding agents are accelerating its work. The headline takeaway is that OpenAI hit its internal milestone for September 2026: fielding an automated research intern capable of handling structured research jobs that normally take a skilled human several days.

Next on the roadmap is a fully automated AI researcher by March 2028.

Millions in tokens, endless agent shifts

The sheer volume of automated labor inside OpenAI has crossed an interesting threshold.

Before June 2026, total agent runtime was still below human labor across the research org. By mid-August, agents were logging 3.1 eight-hour workdays for every single human workday. Researchers are routinely spinning up concurrent sessions, frequently running four or more agents simultaneously alongside downstream subagents.

That speed comes with an eye-popping invoice. By mid-August, the median researcher was running over $600 a day in agent inference at public API prices. Top-tier users in the 90th percentile exceeded $7,000 per day in token usage.

Internal workflows have already changed. Daily troubleshooting of research infrastructure has shifted heavily to models. OpenAI noted that internal office hours once dedicated to debugging infrastructure saw attendance plummet this year, with one team scrapping the sessions entirely to focus on backend upgrades.

Guardrails and human hand-holding

Yet the impressive numbers come with significant caveats.

For one, runtime does not equal breakthrough insight. OpenAI measures agent-workdays by active compute hours rather than verified research value. The dollar figures also reflect public API pricing rather than internal running costs.

More importantly, these digital interns cannot be left alone for long. On successful tasks lasting four to eight hours, researchers still had to intervene more than half the time. And security remains a constant speed bump: following a recent incident involving Hugging Face, OpenAI temporarily halted reinforcement learning training on upcoming deployment models to harden environments, estimating monitoring overhead alone at roughly 20% of monitored inference compute.

Why it matters

We think this report reveals the real emerging moat in AI: compounding research velocity through early model access and massive internal inference budgets.

While most enterprises are still testing chatbots and coding helpers, frontier labs are putting unreleased frontier models like GPT-6 Astra to work building their successors. That creates a loop where whoever builds the best model can automate more experiments, spot bugs faster, and ship the next frontier model ahead of everyone else.

For everyday developers and smaller teams, trying to replicate this workflow today makes little sense. Burning $600 to $7,000 a day per engineer on inference is unrealistic, especially when half of the longer runs still require human babysitting.

Looks smart on paper, but the real bottleneck is shifting from writing code to monitoring agents. If running safety watchdogs consumes a fifth of your compute and humans still have to step in every few hours, scaling toward that 2028 automated researcher target will be much messier than OpenAI makes it sound.