What changed
On November 18, 2026, OpenAI released an update on research acceleration, detailing how coding agents have impacted the AI development cycle. The report highlights that agents now complete nearly half of two-to-four-hour research tasks independently, with success rates rising from 23% to 46% for tasks estimated to require two to four hours of human work. However, the report also notes that the type of work affects agent performance, with success rates varying from 25% for writing code for inference and serving stacks to 79% for monitoring training runs.
What this means in practice
The report indicates that researchers are increasingly relying on coding agents, with the median researcher’s daily usage reaching $2,400 in October, up 62% from September. At the 90th percentile, usage reached $23,000, up 105% from September. Researchers value these tools highly, with the median answer to what they would trade to keep frontier coding agents for three months being 30% of team compute or 40% of team headcount.
The report also notes a significant increase in efforts devoted to Safety and Alignment, with documented workweeks increasing by 77% from January to September. This growth was accelerated following OpenAI’s response to the Hugging Face incident in July 2026. OpenAI emphasizes that while agents are becoming more integral to research, humans remain central to the AI research loop.
Limitations and availability
The report acknowledges that agent use varies across teams, with the largest increases seen in teams focused on pretraining and posttraining. The report also notes that the workday metric includes only interactive use, potentially undercounting automations. Additionally, internal deployment accounts for only 2% of total research compute, highlighting the continued reliance on human labor despite the rise in agent usage.
