The Current

OpenAI reports steep rise in internal 'agent' usage among its researchers

The company's second research-acceleration report quantifies how heavily its own staff now lean on coding agents, while stating humans remain in the research loop.

useful research · for technical · November 19, 2026

On November 18, 2026, OpenAI published an update on 'research acceleration' inside its own organization, building on a September blog post. The report measures how coding agents are being used across its research teams and states that agentic AI usage 'has continued to increase exponentially' since the first report. According to OpenAI, the median researcher now uses around $2,400 USD per day on agentic AI tools at API list prices, up 62% from $1,500 in September, while the 90th-percentile researcher uses around $23,000 per day, up 105% from $11,000. The company defines 'agent workdays' as every eight hours of agent use; these rose from 0.5 to 5.7 per researcher per day, an elevenfold increase, concentrated in pretraining and posttraining teams. By mid-October, OpenAI says more than two-thirds of included researchers averaged a daily peak of at least four concurrent agents, and more than 10% averaged at least sixteen. As of September 2026, the API-equivalent value of agentic AI use was about twice researchers' non-equity compensation, though internal deployment accounted for only 2% of total research compute. OpenAI frames the disclosure as tracking indicators of recursive self-improvement, and says the report should not be read as an intention to pursue full recursive self-improvement.

  • Median researcher uses ~$2,400/day on agentic AI tools at API list prices, up 62% from September
  • 'Agent workdays' per researcher rose elevenfold, from 0.5 to 5.7 per day
  • API-equivalent agent usage was about twice researchers' non-equity pay as of September 2026, but only 2% of total research compute

What it means for you

This is OpenAI reporting on how OpenAI works — a frontier lab describing its own internal R&D, not advice for a normal business. The headline number (researchers running many coding agents at once, spending thousands a day in compute terms) describes an environment almost nothing like a growing organization's day-to-day operations. For most owners and operators, the practical takeaway is near zero: your bottleneck is rarely raw coding throughput, it's unclear processes and handoffs.

Try this

If this report tempts you to picture 'agents running your work,' spend an hour this week writing down one process end to end — who touches it, what gets handed off, where it stalls. That map tells you far more about where you'd benefit from automation than any lab's usage stats.

Who should care

Technical leaders and engineering managers watching how AI-assisted development is evolving, and anyone tracking AI-safety governance and recursive self-improvement disclosures.

Skip this if

You run operations or own a growing business and aren't managing a large software R&D team — this is industry-internal context with nothing to act on.

Sources: OpenAI — read the original

Keep reading

The rest of this story is free with your email.

One field, asked once. It unlocks every story on The Current in this browser, including what it means for you and what to try.

We’ll email occasionally; decline any time. Privacy

← All stories