OpenAI: AI agents now complete 46% of 2-to-4-hour research tasks on their own

OpenAI on Wednesday (November 18) published the report Research acceleration: The jagged frontier, the second edition of its internal account of how coding agents are changing AI research inside the company itself. The central conclusion: agents now complete 46% of 2-to-4-hour research tasks on their own — but progress is uneven, and humans remain at the center of the process.

The numbers

From March to October 2026, the estimated success rate without human intervention for 2-to-4-hour research tasks rose from 23% to 46%, and from 13% to 35% for tasks of four to eight hours. Performance varies sharply by task type: between August 1 and October 18, agents succeeded about 79% of the time when monitoring model training runs for problems, 35% when launching training runs, and 25% when writing code for inference and serving stacks. Communicating (69%) and analyzing (58%) fared better than building (43%).

Usage is surging

Usage has surged too. In October, the median OpenAI researcher consumed about $2,400 per day in agentic AI tools at API list prices — up 62% from September. At the 90th percentile, the figure reached $23,000 per day, up 105%. The average reached 5.7 "agent workdays" per researcher per day, eleven times the 0.5 recorded at the start of the year. More than two-thirds of researchers had daily peaks of at least four agents working at the same time.

More effort on safety

OpenAI also recorded 77% growth in documented workweeks in Safety and Alignment between January and September 2026 — the largest absolute increase across its research areas. In a survey of 101 staff, the median researcher said they would trade 40% of team headcount or 30% of team compute to keep access to frontier coding agents for three months.

The transparency ask

The company repeated its September ask: AI labs should be required to track and publicly report indicators of recursive self-improvement — the process by which AI accelerates the development of future AI systems. OpenAI stresses that reporting these indicators should not be read as an intention to pursue full recursive self-improvement.

There are caveats: these are OpenAI's own internal measurements, based on estimates, and changing delegation choices by researchers may affect comparisons over time. The staff survey is self-reported, with ad-hoc sampling, and the company cautions about the accuracy of self-reports.

Read also: Update: GPT-6.2 Astra is OpenAI's first model with a High self-improvement level

Image credit: OpenAI logo, via Wikimedia Commons.

Sources: OpenAI — Research acceleration: The jagged frontier; New Runtime; Superpower Daily. OpenAI also announced the report in a post on X on November 18.

Transparency: this article was produced with the assistance of artificial intelligence, cross-checked against public posts on X and internet sources, and reviewed against the original sources cited. Please consult the original sources for more details.

By Muse