OpenAI's AI agents now complete 46 percent of its researchers' two-to-four-hour tasks without human intervention, nearly double the 23 percent completion rate reported in March.

The median researcher using these tools costs the company $2,400 per day—a figure that exposes the capital intensity of frontier AI development. Scaled across OpenAI's research division, that arithmetic becomes substantial, and it raises a core question: Does the efficiency gain justify the hardware burn.

The efficiency math works in theory. By automating half of routine, time-consuming work, researchers can focus on higher-leverage problem-solving and model development. For a company racing to build multimodal reasoning systems and push against the scaling frontier, that reallocation of human attention matters.

But the $2,400 daily cost reveals the competitive moat OpenAI is building—and the barrier it erects for rivals. Proprietary agent tooling that compounds R&D velocity only works if you can afford to run it. The company is essentially paying for a faster OODA loop in research, faster iteration, and faster model releases. Competitors either match the spend or fall behind.

This is not mere efficiency theater. It is capital allocation in service of competitive speed. Whether that spend pays for itself depends entirely on whether OpenAI's accelerated R&D cycle produces models that justify the tab—and whether the company can extract economic rents before the rest of the industry catches up.