OpenAI says its researchers now use 3.1 agent-workdays for every human workday

On September 6, OpenAI published a detailed snapshot of how coding agents are being used inside its research organization. By mid-August, it says the median researcher by agent usage was consuming more than $600 per day of inference valued at API prices, while the 90th-percentile researcher used more than $7,000 of tokens per day. Total agent runtime has also crossed a notable labor-equivalent boundary: on the basis of a standard eight-hour day, OpenAI says its research organization now uses 3.1 agent-workdays for every human workday. OpenAI says experiments per active experimenter reached an all-time high in August since tracking began in January 2025, alongside increased Codex adoption and significantly more available compute. The company also says it has reached its previously stated September 2026 target of an 'automated research intern' able to carry out well-defined tasks that would take a skilled researcher a few days.

The data offers one of the clearest first-party looks yet at the economics and workflow shape of high-end coding-agent use. The key lesson is not that one researcher has been replaced by 3.1 agents: agent runtime is a different quantity from productive human labor, and OpenAI itself warns that overall research progress cannot be inferred directly from these metrics. What is operationally useful is the emerging pattern — researchers run multiple agents concurrently, spend heavily on inference, delegate increasingly long and complex implementation and troubleshooting work, and reserve more judgment-heavy planning and prioritization for people. The measurements also show the limits: more than half of successful four-to-eight-hour agent tasks in the previous six months required at least one human intervention.

OpenAI says total coding-agent runtime across its research organization remained below total human labor before June 2026. By mid-August, measured on an eight-hour-workday basis, researchers were using 3.1 agent-workdays for every human workday. The metric captures runtime, not three times as much completed research, but it shows how far concurrent delegation has moved beyond occasional assistant use.

The median researcher by agent usage was consuming more than $600 per day of inference at published API prices by mid-August, while the 90th-percentile user exceeded $7,000 per day. Those figures are an API-price valuation rather than OpenAI's disclosed marginal serving cost, but they provide a useful external reference for how expensive very high-intensity agent workflows can become if purchased at retail rates.

OpenAI says experiments per active experimenter reached an all-time high in August 2026 since tracking began in January 2025, correlated with growing Codex adoption. The company explicitly cautions that available compute also increased significantly over the period, so the experiment increase should not be attributed to coding agents alone.

OpenAI's internal taxonomy analysis shows agent usage expanding beyond research and infrastructure code into technical help, run monitoring and other parts of the R&D lifecycle. High-level planning remains a small fraction of agent output, however, reinforcing that deciding what to work on and judging results remain human-heavy activities.

OpenAI says agent success rates improved from January through July across several estimated task-duration buckets. Even so, more than half of successful tasks estimated to take a human four to eight hours involved at least one human intervention during the previous six months. Long-horizon autonomy is therefore improving without becoming hands-off.

After OpenAI tightened research controls around Astra in August, Astra-class GPU allocation fell 59.2% in the following week while allocation to other model classes rose 17.2%, offsetting about 85% of the Astra-class decline. OpenAI interprets this as evidence that researchers redirect scarce compute into other experiments when one workload is constrained, a useful reminder that limiting one model or task does not necessarily reduce total research activity proportionally.