openai.com web signal

OpenAI Says Coding Agents Finish 46% of 2-4 Hour Research Tasks

OpenAI Agents Coding Tools ai-business

TL;DR

  • OpenAI reports coding agents now complete 46% of its two-to-four hour internal research tasks without human help, up from 23% in March.
  • Capability is uneven: 79% on monitoring model-training runs but only 25% on writing inference and serving-stack code.
  • Median researcher daily agent use hit $2,400 in October, a 62% jump from September; 90th-percentile daily use reached $23,000.

OpenAI says coding agents now finish 46% of its two-to-four hour research tasks without human help, up from 23% in March, in a company post titled "Research acceleration: The jagged frontier" published November 18. Four-to-eight hour tasks moved from 13% to 35% over the same window.

The picture is uneven, as the title warns. Measured between August 1 and October 18, agents landed 79% on monitoring model-training runs, 35% on launching training runs, and 25% on writing inference and serving-stack code. "The agents are not taking over AI research," the post concludes. "They are taking over specific slices."

Internal spend is climbing fast. OpenAI reports median researcher daily agent use valued at $2,400 in October, a 62% jump from September, with 90th-percentile daily use hitting $23,000. The company writes that "more than two-thirds of researchers averaged a daily peak of at least four agents," and "more than 10% averaged a peak of at least sixteen."

A survey of 101 research staff posed a hypothetical: if your team lost access to coding agents better than GPT-5.2 for three months, how much would you trade to get them back? The median answers were 30% of team compute or 40% of team headcount. Safety and Alignment showed the largest absolute increase in documented researcher workweeks of any research area, up 77% from January to September.