OpenAI Says Its Team Relies on AI Coding Agents For More Work Hours Everyday Than Human Researchers

OpenAI has published a detailed self-assessment of how AI coding agents are changing the pace and shape of research inside the company, revealing that its research staff now collectively rely on agents for more work-hours than they put in themselves.

According to the company, as of mid-August, its research organisation was drawing on the equivalent of 3.1 agent-workdays - measured against a standard eight-hour shift - for every workday logged by a human researcher, a threshold the company says it crossed only since June.

Usage climbing sharply across the board

OpenAI said that at the start of the year, the median researcher tracked by internal usage metrics was using coding agents only sparingly. By mid-August, that same median researcher was spending more than $600 a day on inference at API pricing, integrating agents into daily work, while the heaviest users - the top 10th percentile - were running through more than $7,000 worth of tokens daily. The company noted that this growth in usage is outpacing every other function at OpenAI, with output for the typical research employee rising 124-fold since December.

The trend extends to how many researchers are running multiple agents side by side. OpenAI said the share of researchers juggling four or more concurrent agent sessions - counting both agents launched directly and further sub-agents they spawn has been steadily climbing through the year.

More code, more experiments

OpenAI framed research as a chain of steps - designing improvements, building evaluation tools, testing at scale, catching errors, and folding successful ideas into training runs - and said agents are visibly speeding up several links in that chain. Company-wide, the amount of code shipped per active contributor has risen several-fold compared with pre-2025 levels, and the number of experiments run per active experimenter hit an all-time high in August, a trend OpenAI links to wider adoption of its Codex coding tool.

The company was careful to flag that these figures are easier to collect than to interpret, since a rising line count does not necessarily map cleanly onto genuine research progress, and cautioned that compute availability, not just agent capability, may increasingly determine how fast experiments can run.

Agents taking on bigger, longer jobs

Drawing on a research-classification framework built by Epoch AI, OpenAI mapped the kind of work its agents are being asked to do - spanning deciding priorities, designing experiments, building code, running training jobs, analysing results and communicating findings. It found growth across every one of those categories since January, though writing research and infrastructure code remains by far the largest use case. The company also pointed to a decline in traffic on an internal help channel where researchers used to seek troubleshooting support from colleagues, saying agents appear to be absorbing much of that debugging work instead, with one team reportedly shutting its human-staffed support sessions altogether.

OpenAI's internal tracking also suggests agents are succeeding more often on harder, longer-duration tasks than they did at the start of the year — though the company noted that jobs estimated to take a human four to eight hours still needed a human to step in and intervene more than half the time.

Published on: Monday, September 07, 2026, 10:12 AM IST