Money and business in the Middle East.

AI

OpenAI says AI agents doubled success on multi-hour research tasks

OpenAI reports agent success on 2-to-4-hour research tasks rose from 23% to 46%

· Last verified: 20 Nov 2026 (OpenAI)

Summary

  • OpenAI reports agent success on 2-to-4-hour research tasks rose from 23% to 46%
  • Median researcher agent spend hit $2,400 a day in October, up 62% from September
  • First published internal measure of how far agents have entered a frontier lab's own research

The latest

Estimated agent success on research tasks lasting two to four hours climbed from 23% in March to 46% in October, OpenAI said in a report titled Research acceleration: The jagged frontier. Tasks running four to eight hours rose from 13% to 35%. The lab attached explicit limits to the data, saying higher activity metrics do not necessarily mean faster research progress.

Details

  • The headline numbers: Agent success on two-to-four-hour research tasks doubled from 23% to 46% between March and October, the report said, while four-to-eight-hour tasks moved from 13% to 35%. The figures estimate agent performance inside OpenAI's own research workflow rather than public benchmark results.
  • Task-level breakdown: Self-reported success varied sharply by job type: 79% for monitoring training runs, 35% for launching them, and 25% for writing code for inference and serving stacks. The spread is the basis for the jagged frontier framing, in which capability is uneven across research work.
  • Spending: Measured at published API prices, the median researcher's daily agent usage reached $2,400 in October, up 62% from $1,500 in September. The 90th percentile hit $23,000, a 105% rise from $11,000. Through September 2026, the API-equivalent value of agent use was roughly twice non-equity compensation.
  • Scale of use: Agent workdays per researcher rose from 0.5 to 5.7, an eleven-fold increase counting every eight hours of agent use as one workday. By mid-October, more than two-thirds of researchers covered averaged daily peaks of at least four concurrent agents, and more than 10% reached at least sixteen.
  • Unassisted performance: Between 1 August and 18 October 2026, success without human intervention was 69% for communication tasks, 58% for analysis and 43% for building. Answering technical questions succeeded 72% of the time, while building evaluation, reinforcement learning and agent environments succeeded 29%.
  • Output measures: Documented work weeks on the safety and alignment team rose 77% between January and September. Logged experiments per active researcher ran at 1.68 times the 2025 baseline in the four weeks ending 18 October. Internal deployment accounts for 2% of total research compute.
  • The capability trend: The report put OpenAI's ECI frontier at plus 16.4 points a year since August 2025, against an Epoch trend of 14 points a year since late 2024. One safety researcher working on reinforcement learning environments said: "The model generally can't one-shot come up with well-posed problems that it cannot solve itself."
  • The survey: A poll of 101 research staff conducted from late August to early September 2026 found the median respondent willing to give up 40% of team headcount or 30% of team compute to keep advanced coding agents for three months. Asked about January 2027, the median rose to 50% of compute.
  • Method and caveats: OpenAI said its measure generally captures interactive use only and may therefore understate automation. Time allocation was estimated from internal Slack posts, merge requests and documents from January to September 2026. Survey responses are self-reported and the sample is non-systematic.
  • The footnote: The report disclosed an internal change on 21 August that lowered measured token spending until it was reversed on 10 September, a gap affecting the September baseline against which October's spending increases are calculated.

Between the lines

The gap between 79% success on monitoring training runs and 25% on writing inference and serving code is the report's central finding, not the doubling. Agents cluster in supervision and communication work and fall off in open-ended construction. That the measure counts interactive use only, as the report concedes, means the published figures describe a floor rather than a ceiling.

What's next

Whether OpenAI repeats the measurement past October, and whether the survey's January 2027 question is actually asked, would test the report's own caution that rising activity metrics do not equal faster research progress.

Source: OpenAI

OpenAI says AI agents doubled success on multi-hour research tasks · INXEN