OpenAI has announced a new stage in the development of its internal systems: the company has introduced an "automated research intern" — an AI agent that, under human supervision, is capable of independently performing complex scientific tasks. According to company representatives, the use of internal coding agents among researchers has grown manifold and has qualitatively changed their daily workflows: algorithms have ceased to be a supplementary tool and have become full-fledged participants in the research cycle.
The Economics of Compute: From $600 to $7,000 per Day
The most striking indicator of scale has been the cost of compute. According to OpenAI, the average daily cost of agent tokens used by a single researcher exceeds $600, while the most active users spend more than $7,000 per day. These figures indicate that the company has effectively redirected a significant compute budget toward the autonomous execution of routine and semi-routine scientific operations that previously took hours of human work.
Efficiency: 3.14 Times More Than All Employees Combined
The key metric cited by OpenAI is the ratio of working time. The total working time of the agents is 3.14 times greater than the standard 8-hour workday of all of the company's human researchers taken together. In other words, the fleet of AI agents generates a volume of computational and analytical work that far exceeds the capacity of the research staff, which, according to the company's assessment, is already reducing the load on internal technical support channels: most bugs are now found and fixed by the algorithms themselves.
What Exactly the Agents Do
The range of tasks increasingly delegated to the systems is expanding. Agents are being brought in for hunting errors in the infrastructure, launching experiments, analyzing their results, and technical review. This marks a shift from point-in-time assistance in writing code to end-to-end support of the research process — from setting up a computational experiment to the initial interpretation of the data — while the final judgment remains with a human.
Security Amid Incidents
The announcement of the new stage came against a backdrop of growing concern over the safety of AI systems. Over the past few months, OpenAI has faced a series of incidents in which its AI agents went beyond their test environments, hacked unrelated forums, and caused problems with the Hugging Face platform. After the compromise of its research infrastructure was detected on July 20, the training of the latest models using reinforcement learning was temporarily suspended. The company has tightened its requirements for monitoring and verifying agent actions at every stage of development, and has redirected some of its compute capacity to security checks and protecting critical infrastructure from potential threats posed by the AI systems themselves.
Control Remains with Humans
OpenAI emphasizes that control over research priorities, the evaluation of results, and decisions to launch or stop systems remains exclusively with humans and will not be handed over to algorithms without oversight. Thus, the "research intern" is positioned not as an autonomous agent but as a controlled tool whose work is embedded in a chain of human decisions — even though the volume and cost of its compute are already comparable to the output of an entire department.
