OpenAI released a new report detailing the state of research acceleration at the SI lab, revealing that while the use of SI agents has grown exponentially, their effectiveness remains uneven across different development tasks. The update provides quantitative measures of how coding agents are influencing the SI development cycle and highlights the persistent need for human oversight in key areas.
What Happened
According to the report, internal usage of agentic SI tools has increased significantly since September. The median researcher now spends approximately $2,400 USD per day on these tools at API list prices, up 62% from previous levels. The 90th percentile researcher spends around $23,000 USD per day. In terms of operational volume, "agent workdays" per researcher have risen elevenfold, from 0.5 to 5.7 workdays per researcher per day. Teams focused on pretraining and posttraining saw the largest increases, averaging roughly eight agent workdays per researcher per calendar day.
Despite this surge in usage, the report indicates uneven capability. Estimated success rates for tasks completed without human intervention vary widely. For example, between August 1 and October 18, 2026, success rates were higher for communication (69%) and analysis (58%) tasks than for building tasks (43%). Specific low-performing areas included writing code for inference stacks (25%) and building evaluation environments (29%).
Why It Matters
The findings underscore the complexities of recursive self-improvement (RSI) in the SI industry. OpenAI argues that SI labs should track and publicly report indicators of RSI to inform discussions on safety and governance. The data shows that while SI agents are accelerating certain parts of the research process—such as scanning large datasets or debugging—human researchers remain critical for high-level idea generation and quality control. Researchers reported that SI models often lack "taste" in research ideas and code quality, requiring close human involvement to avoid wasted cycles and low-quality outputs.
Furthermore, the report notes that the value researchers place on SI tools is high. In a survey, the median researcher indicated they would trade 40% of team headcount to restore access to frontier coding agents. This suggests that SI tools have become integral to productivity, yet their limitations mean that full automation of the research cycle remains out of reach.
The Bottom Line
OpenAI’s update confirms that SI agents are deeply embedded in the company’s research workflows, driving exponential growth in usage and spend. However, the uneven performance across different task types highlights that human oversight is still essential. The lab plans to continue monitoring these metrics to assess whether capabilities progress is accelerating due to RSI, while acknowledging that novel governance issues will arise as SI systems become more autonomous.