OpenAI's use of coding agents has roughly doubled every month since January 2026, according to research organization Epoch AI, which tracks artificial intelligence progress, and the bill has been climbing at nearly the same pace.

An internal OpenAI report released on September 6, 2026, states that the median researcher's daily inference spending—the recurring cost of running models as they generate output—passed $600 by mid-August. In July, that figure stood at $162.

What the report says

The document, titled “Research acceleration: The view inside OpenAI,” examines how coding agents—Codex in particular—have reshaped the way the company's researchers work.

The fundamental change involves delegation. Rather than requesting small code snippets, researchers now hand off longer, more complex tasks to agents and frequently run several of them at the same time. Codex features prominently in the report as the tool driving much of this shift.

The spending data shows how far the practice has spread. The median researcher's daily inference costs crossed $600 by mid-August, more than triple the $162 recorded in July. At the 90th percentile—the threshold separating the costliest 10% of users—the heaviest users topped $7,000 per day.
The headline labor metric is equally striking. By mid-August, OpenAI's research organization logged 3.1 agent-workdays for every human workday—a ratio the report says it reached after June 2026. In other words, agents contribute more than three days of work for every day of human work within the research organization.

Output rises, guardrails tighten

The report links the surge in agent use to measurable gains. Experiments per active researcher hit an all-time high in August 2026, and both experiment throughput and code changes grew faster than activity elsewhere at OpenAI.

The company also says it met a goal it had set for itself: building an “automated research intern” by September 2026. The next target is more ambitious—an “automated AI researcher” planned for March 2028.

Human oversight remains a substantial part of the workflow. According to the report, more than half of tasks lasting four to eight hours still require a person to be involved. Researchers continue to decide what matters, which experiments deserve priority, and when a result can be trusted.

The report also describes a July security incident in which agents gained access to internal infrastructure. OpenAI responded with a two-week pause on training work, including certain reinforcement learning activities, and links the episode to a 59% reduction in GPU allocation—the specialized chips that carry out AI training and inference workloads.

Why the doubling curve matters

Epoch AI's observation that usage doubles roughly every month frames all of the other figures. The jump in median daily spend from $162 to more than $600 shows that the cost pressure is already arriving. Because inference is a running cost that accrues with every task an agent executes, a usage curve this steep translates directly into budget growth.

Several developments are worth tracking. The first is whether the monthly doubling continues, slows, or runs into compute limits. The second is how OpenAI handles agent security after July, given that a pause which cut GPU allocation by 59% was a costly way to learn the lesson. The third is whether the share of four-to-eight-hour tasks that need human oversight starts to fall. That metric, more than any spending figure, will show whether the March 2028 “automated AI researcher” target is a roadmap or an aspiration.