OpenAI’s research agents are improving at longer tasks, but human oversight remains essential. Figures show higher rates for assignments that would take people several hours, while highlighting why greater AI use does not automatically mean higher productivity.
In a November 18 update, OpenAI said agents independently completed an estimated 46% of tasks requiring two to four hours of human work in October, up from 23% in March. For tasks estimated at four to eight hours, success rose to 35%, compared with 13% in March.
These figures estimate the time a person would need, not how long an agent works. Results vary by assignment. Between August 1 and October 18, agents had an estimated success rate of 79% for monitoring training runs, but 25% for writing code used to run models and serve responses.
Researchers continue to set goals, review agent output and make corrections rather than handing over projects. OpenAI cautioned that researchers may be changing the tasks they delegate as the technology improves, influencing reported success rates.
The company is tracking agent usage, code changes, experiments and model capabilities. Median daily agent use by researchers was valued at around USD 2,400 at API list prices in October. This reflects usage, not spending.
Also Read: Arizona Investigates OpenAI Over Alleged Rogue AI Bank Infiltration
OpenAI said spending, code and experiment counts cannot measure productivity. More activity may not produce better findings, particularly when human review and result quality are considered. The company plans to refine its methods as it studies how agents affect research progress.
The findings point to a shift in research: agents can take on execution, while people remain responsible for deciding what matters and whether results are reliable. The gains suggest improving capabilities, not a process that runs without human direction.