OpenAI Safety has published insights into the internal deployment of coding agents, detailing how these systems are influencing AI research within the organization. The report outlines early data on agent usage, experiment velocity, task complexity, and the overall acceleration of research progress.
Key Points
- OpenAI reached its goal of having an automated research intern by September of this year, a system capable of carrying out well-defined research tasks under human direction that would typically take a skilled researcher several days.
- The organization is making progress toward creating an automated AI researcher by March 2028.
- By mid-August, the median OpenAI researcher was using coding agents daily, incurring more than $600 per day in inference costs at API prices.
- The 90th percentile user in the research organization was utilizing more than $7,000 of tokens per day by mid-August.
- As of mid-August, the research organization uses 3.1 agent-workdays of effort for every workday of human labor.
- Following a recent incident, OpenAI paused reinforcement learning (RL) training on its latest models intended for deployment to harden research environments and expand monitoring systems.
Context
According to OpenAI Safety, the daily work of OpenAI researchers has substantially changed over the course of this year, with coding agents being used throughout the day, often in concurrent sessions. Total agent usage is rapidly increasing, outpacing growth among other OpenAI teams. Researchers are contributing code faster and running more experiments, with agents handling increasingly complex tasks and succeeding more often.
Why It Matters
This transparency from OpenAI provides insight into the practical application of AI agents in accelerating research within a frontier lab. It offers a view into how internal development practices are evolving, which can inform other labs and deployers about the potential for agentic systems to impact research velocity and operational costs. The disclosure also highlights the ongoing commitment to safety measures and the iterative process of balancing capability advancements with risk mitigation.
What To Do
- Note the stated timeline for an automated research intern by September of this year and an automated AI researcher by March 2028.
- Observe the reported daily inference costs for median and 90th percentile researchers to understand potential resource allocation for agent-driven research.
- Watch for further disclosures from OpenAI regarding their progress toward Research Superintelligence (RSI) and evolving transparency approaches.
- Consider the implications of agent-workday ratios on research productivity and resource planning.