Are AI agents now doing more research work than humans?

OpenAI announced in September 2026 that it had reached a goal set "last fall" of having an automated research intern by September 2026 — a system able to "carry out well-defined research tasks under human direction, including work that would take a skilled researcher several days." It backs the claim with internal measurement rather than a product description alone: "the research organization logs 3.1 agent-workdays of effort for every eight hours of human labor," a ratio that was still below total human labor as recently as June 2026. OpenAI also reports daily inference spend at API prices reaching more than $600 for the median researcher using coding agents by mid-August, and more than $7,000 at the 90th percentile, alongside a growing number of researchers running "four or more agents simultaneously."

OpenAI frames the acceleration as agent contribution to specific research steps — "developing ideas, creating tests to measure results, building systems to run experiments at scale, identifying bugs and safety problems, and incorporating successful changes into model training" — rather than agents setting direction. Using a six-phase framework developed by Epoch AI, OpenAI classified coding-agent activity across research stages and found rising agent use in implementation and experimentation between January and August 2026, while "high-level planning remained rare." Measured success rates "generally increased" across task-difficulty categories between January and July, though more than half of successful tasks expected to take a person four to eight hours still required at least one human intervention — OpenAI's own evidence that the acceleration is uneven across task type rather than uniform.

This supplies a concrete instance of what How fast is AI accelerating its own development inside labs? proposed in the abstract: a lab disclosing verifiable AI-R&D-pace metrics. OpenAI's headline number is differently shaped — a ratio of agent-workdays to human-labor-hours crossing above 1:1, rather than a percentage share of total R&D work — but it is the same move, publishing internal activity data for public scrutiny. It also supplies the measured-activity half of Can global standards pace frontier AI as much as alignment research?, which stated OpenAI's policy position on pacing RSI; here OpenAI publishes the rising agent-workdays, rising inference cost, and rising success rates it says should inform that debate. It sits apart from Could automated AI research compress years of progress into months?: that note is a participant's projection about a future feedback loop, while this source reports what OpenAI says its own research organization is measuring today, not a forecast of what automation could eventually compress.

The excerpt gives a ratio of agent-workdays to human-hours, not a ratio of research output or research quality — more agent-hours logged is not the same as proportionally more research progress, especially since OpenAI's own account shows planning and judgment remain human-led and difficult tasks still require frequent intervention. The data is also self-reported by the lab whose stated goal was exactly this milestone, and the excerpt describes no independent verification. The implication, at the strength the evidence allows, is that labor substitution in AI research is underway and accelerating by OpenAI's own measure, but the step from "agent-workdays logged" to "research progress achieved" is not something this excerpt measures.