Over the past year, AI systems have become increasingly adept at contributing to the research and development process, potentially transforming the way AI is developed and deployed. Today, OpenAI is releasing an update on the progress of research acceleration within the organization, based on internal measurements of how coding agents have influenced the development cycle.

This report builds on OpenAI’s previous research acceleration report from September, adding new quantitative and qualitative measures to better understand the role of AI tools in the research process. The findings highlight a significant shift in how researchers are utilizing coding agents, with a notable increase in their ability to complete tasks autonomously.

According to the report, coding agents are now completing 46% of researchers’ 2 to 4 hour tasks on their own, compared to just 23% in March. The median researcher at OpenAI now spends approximately $2,400 per day on agentic AI tools, up 62% from $1,500 in September. The 90th percentile researcher spends around $23,000 per day, a 105% increase from $11,000. This trend reflects both increased usage and shifts to more expensive models.

The report defines an “agent workday” as every eight hours of agent use, providing a more accurate measure of “wall clock time” than financial expenditure. Overall, agent workdays per researcher have increased from 0.5 to 5.7, a 1100% increase. This increase is particularly pronounced in teams focused on pretraining and posttraining, where agent runtime averages about eight workdays per researcher per calendar day.

OpenAI also observed a significant increase in the use of concurrent agents. By mid October, more than two thirds of researchers averaged at least four agents working concurrently, with more than 10% averaging at least sixteen concurrent agents.

To better understand the allocation of resources, the report compares monthly API equivalent spending on agentic AI use with monthly non equity compensation for human researchers. As of September 2026, the API equivalent value of agentic AI use in OpenAI’s research organization was about twice its non equity compensation, with higher ratios in pretraining and posttraining than in other parts of the research organization. Despite this, internal deployment, which includes internal Codex, ChatGPT, and API usage, accounts for only 2% of total research compute.

Researchers are also increasingly allocating their time across different research areas. Using a method that divides researcher weeks equally across verified projects, the report finds that the amount of time dedicated to each area has grown, with Safety & Alignment seeing the largest absolute increase. This research area covers workstreams such as alignment evaluations, monitorability, and reward hacking research. A notable inflection point in the growth of Safety & Alignment coincides with the aftermath of the Hugging Face incident and OpenAI’s subsequent efforts to develop and improve safety cases for frontier AI training.

To examine how agent use affects research productivity, the report provides both quantitative and qualitative measures. For example, using an updated methodology, the analysis found that agents have become more likely to complete tasks successfully without human intervention, including tasks estimated to require several hours of human work. This trend has generally continued, with changes reflecting both capability improvements and researchers’ changing choices about when and how to use agents.

The report also details how often researchers use coding agents for individual tasks, with higher weight placed on short tasks such as answering technical questions compared to token intensive tasks such as launching and monitoring a training run or implementing a complex feature.

Through surveys and interviews, researchers reported valuing coding agents significantly. A survey conducted from late August to early September 2026 found that the median researcher was willing to trade $55,000 to restore access to frontier coding agents. Additionally, 27 out of 28 respondents in a late October survey chose or leaned toward the latest internal model over a combination of a mid career human research engineer and a 2024 model.

In conclusion, the report underscores the growing importance of AI coding agents in the research process, with significant changes in both usage and outcomes. OpenAI continues to emphasize the need for transparency in tracking and reporting indicators of recursive self improvement (RSI) to inform discussions on the capabilities, risks, and required safeguards for increasingly powerful AI.

Source: https://openai.com/index/research-acceleration-jagged-frontier

Thinking about building an AI product?

Get in Touch