The cost of AI usage by OpenAI researchers and the scale of agent deployment have risen sharply in 2026. [Photo: Shutterstock]

[DigitalToday reporter Choi Jae-won (최재원)] Use of AI agents by OpenAI researchers has surged, rapidly changing how AI research and development is conducted. AI is carrying out more tasks without human involvement, but an analysis said human researchers still play an important role in finding key ideas and verifying results.

On Nov. 19 local time, Japanese media outlet Gigazine reported that OpenAI released a report a day earlier titled "Research Acceleration: The Jagged Frontier" and disclosed the status of AI agent use among its internal researchers.

As of October, the median daily AI usage per researcher, converted at the API list price, was $2,400 (about 3,196,300 won). For the top 10 percent, it reached $23,000 (about 30,631,400 won). Compared with September, those figures were up 62 percent and 105 percent, respectively. The figures reflect AI usage converted into prices, not actual spending.

Time spent using agents also rose sharply. Average agent workdays per researcher per day increased about 11-fold, to 5.7 days from 0.5 days. An agent workday refers to 8 hours of AI operation.

The practice of using multiple agents simultaneously also spread. As of mid-October, more than two-thirds of researchers were found to run at least 4 agents simultaneously in a day.

AI's ability to carry out tasks independently also improved. The share of research tasks expected to take humans 2 to 4 hours that were completed without human intervention rose to 46 percent in October from 23 percent in March. The completion rate for 4 to 8-hour tasks also increased, to 35 percent from 13 percent.

Performance differences by task type were large. Autonomous completion was 69 percent for communication-related tasks and 58 percent for analysis, but system building was limited to 43 percent. Reasoning and writing code for service systems was just 25 percent.

Researchers' reliance on AI also increased. In a survey of 101 researchers, the median response said that if they could not use state-of-the-art coding agents for three months, they would be willing to give up 40 percent of team headcount or 30 percent of computing resources to get them back.

Still, OpenAI acknowledged major flaws remain, including inaccurate code generated by AI, repetitive data and incorrect analytical assumptions.

OpenAI presented the results as material to identify signs of "Recursive Self-Improvement" (RSI), in which AI accelerates the development of next-generation AI. It stressed that AI has not reached a stage of repeatedly developing and improving performance on its own without humans, and that human oversight is needed across research.

Keyword

#OpenAI #Research Acceleration: The Jagged Frontier #Recursive Self-Improvement #Gigazine #AI agents
Copyright © DigitalToday. All rights reserved. Unauthorized reproduction and redistribution are prohibited.