OpenAI researchers' use of artificial intelligence has increased sharply. The median AI usage by researchers exceeded $600 a day, or about 800,000 won, while some researchers were tallied as doing more than $7,000 a day, or about 9.4 million won, in AI work. Another analysis said AI agents have entered a phase in which they put in more working time than humans.

On Sept. 7, local time, online media outlet Gigazine reported that OpenAI on Sept. 6 released a research report detailing how its researchers use AI and how usage has changed. The report shows how deeply AI is being used in actual research tasks at AI development sites.

The most striking change is researchers' AI usage. OpenAI researchers' output token usage increased 124-fold from November 2025 to August 2026. Converted using consumer API pricing, the median daily usage as of August exceeded $600. For the top 10 percent of researchers, daily AI usage topped $7,000.

How AI is used is also moving beyond simple question-and-answer. As of August 2026, about 70 percent of researchers were found to be operating at least four AI agents simultaneously. That means assigning tasks to multiple AI agents and conducting research in parallel has spread quickly in research settings.

The amount of work time AI agents put in stands out further. According to OpenAI, in June 2026 AI agents' labor time surpassed humans' labor time, and by mid-August it had increased to 3.14 times the human level. AI moved beyond assisting researchers and overtook humans in actual time spent on work.

Code output also rose sharply. In July 2026, engineers' code output increased 7.02-fold compared with the pre-2025 average. Researchers' areas of AI use also expanded around actual research execution and operational tasks, including code research and composition, technical assistance and review, open monitoring and debugging.

But AI has not fully automated all research tasks. Of tasks that require 64 to 128 hours by humans, only 16 percent were completed by AI from start to finish without human intervention. Some 51 percent required at least one instance of human intervention, 26 percent failed, and 7 percent had tool errors. It suggests there are still limits to fully autonomous execution as tasks take longer.

OpenAI also disclosed some indications of internal use of research AI models. The company said GPT-6 Astra was being used for internal research before its official announcement. Training activity targeting Astra-series models was also confirmed through reinforcement learning trends, but related training was found to have declined after the Hugging Face attack.

The report is seen as material showing how far OpenAI's push toward a phase in which "AI conducts AI research" has progressed. OpenAI has previously presented a goal of implementing a "truly automated AI researcher" by March 2028.

In the end, what stands out in these data is not simply that AI usage has increased. It is that researchers use multiple AI agents at once, AI puts in more working time than humans, and code output is being lifted sharply. But limits remain in fully automating long-duration tasks, and a key focus is expected to be how much AI agents can expand the scope of work they can do without human intervention.