OpenAI Reports Uneven Progress in AI-Driven Research Acceleration
OpenAI has released an update on the state of research acceleration within its operations, highlighting how coding agents are reshaping the AI development process. The report, published on November 18, 2026, details both the progress and limitations of these agents, as well as their impact on human involvement in AI research.
Agents' Performance and Task Success Rates
According to OpenAI, coding agents have demonstrated significant improvements in completing research tasks. From March to October 2026, the success rate for tasks requiring two to four hours of human effort rose from 23% to 46%. For more complex tasks requiring four to eight hours, the success rate increased from 13% to 35%. However, success rates vary widely depending on the nature of the task. For instance, agents achieved a 79% success rate in monitoring training runs but only 25% when tasked with writing code for inference and serving stacks. This uneven progress underscores the "jagged frontier" of AI capabilities, where performance remains task-dependent and humans continue to play a critical role in the research loop.
Rising Usage Among Researchers
The report also highlights a surge in the use of coding agents among researchers. By October 2026, the median daily expenditure on these tools reached $2,400, marking a 62% increase from September. At the 90th percentile, daily usage costs climbed to $23,000, a 105% jump over the same period. A survey conducted by OpenAI revealed that researchers place high value on these agents, with median responses indicating they would trade 30% of team compute or 40% of team headcount to retain access to frontier coding agents for three months.
Focus on Safety and Alignment
OpenAI reports a notable shift in research priorities, with increasing human effort being directed toward Safety and Alignment. Documented workweeks in this area grew by 77% between January and September 2026, driven by both new hires and a reallocation of existing researchers. This focus intensified following OpenAI's response to the Hugging Face incident in July 2026, emphasizing the organization's commitment to addressing safety and governance challenges as AI capabilities advance.
Implications for Recursive Self-Improvement
The findings contribute to OpenAI's broader efforts to track and report indicators of recursive self-improvement (RSI), a process where AI systems accelerate the development of future AI technologies. OpenAI stresses the importance of transparency in monitoring RSI as a means to ensure human oversight and safe scaling of AI systems. The report concludes that while AI-driven research acceleration holds immense potential, it also introduces novel technical and governance challenges that require careful management.