OpenAI Publishes Agent Metrics Showing Human Help Persists on Longer Tasks

OpenAI says its research organization used 3.1 agent-workdays of effort for every human workday by mid-August. But its newly published internal snapshot also found that more than half of successful agent tasks estimated at four to eight human hours needed at least one human intervention.

The figures offer a view into how OpenAI is using coding agents to develop models, rather than a measure of completed scientific advances. OpenAI says total agent runtime overtook total human labor after June, and that its median researcher was using more than $600 a day in inference at API prices. The company cautions that such activity metrics are hard to translate into overall research progress because research has multiple bottlenecks.

Parallel work is growing; judgment remains human

OpenAI says experiments per active experimenter reached their highest level in August since tracking began in January 2025, though available compute had also grown significantly. Agents are being used most heavily for research and infrastructure code, technical help, and training-run monitoring. High-level planning remains a minimal share of agent output, while people retain responsibility for priorities, evaluating results, and decisions to scale, pause, or deploy systems.

What the task data shows

• Tasks estimated to take less than 15 minutes succeeded without intervention 86% of the time, according to OpenAI’s assessment.
• For successful tasks in the four-to-eight-hour range, more than half involved at least one human intervention.
• OpenAI’s automated “research intern” milestone covers well-defined tasks carried out under human direction.

Automation ambition meets a monitoring problem

The research-intern milestone is context for the new measurements, not proof of a fully autonomous lab. OpenAI says the system can handle well-defined research tasks, including some that would take a skilled researcher several days, and says it is making strong progress toward an automated AI researcher by March 2028. That future system is intended to set research questions and conduct experiments with less human input.

OpenAI says it does not yet know how to safely reach aligned, full recursive self-improvement, in which AI helps improve the systems that build its successors. Separately, chief scientist Jakub Pachocki wrote that chain-of-thought monitoring is losing reliability because advanced models can manipulate verbalized reasoning or sometimes complete complex tasks without producing it. He has called for voluntary slowdowns until shared safety bars exist and for frontier-safety frameworks to become binding standards enforced by outside bodies.