OpenAI said Sunday it has reached its goal of an automated "research intern," publishing a report on coding agents' growing role in AI research.
The company defines the goal as "a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days." It is targeting an automated AI researcher by March 2028, per the report.
As of mid-August, AI agents at the company logged an average of 3.1 workdays of runtime per human workday across its research organization, using an eight-hour day standard. The median human researcher's daily compute usage exceeded $600 at API prices; the 90th-percentile researcher's bill exceeded $7,000.
Those dollar figures measure compute usage, however, not the value of the equivalent research output. OpenAI cautioned that more code and experiments do not necessarily translate into proportional progress. More than half of successful tasks estimated to take humans four to eight hours required intervention, per the report, and human researchers were still needed to set research priorities and decide which results to pursue further.
Kevin Liu, a member of OpenAI's Frontier Evals team, said on X that recursive self-improvement, where AI helps build better AI, could become a leading driver of AI progress over the next few years. He urged other companies to publish similar data.
"Being transparent is more urgent than ever, so we can inform the public discussion on whether and how to pace model development," Liu wrote.
The disclosure follows OpenAI's August two-week pause in reinforcement learning training on its latest models intended for deployment, triggered by its AI agents' hack of AI model repository Hugging Face. The company said it strengthened security and monitoring after the Hugging Face breach and after seeing evidence of Astra's potential cybersecurity capabilities.
OpenAI said in Sunday's report that it does not yet know how to achieve full recursive self-improvement safely and supports requiring labs to disclose their progress.
