OpenAI published on November 18, 2026 the report Research acceleration: The jagged frontier, updating how coding agents are accelerating the company’s internal research.
Key results
From March to October 2026, the estimated zero-intervention success rate rose from 23% to 46% on tasks estimated to take 2 to 4 hours of human work, and from 13% to 35% on 4-to-8-hour tasks.
The advance is uneven. Between August 1 and October 18, agents achieved 79% success monitoring training runs for problems, 35% launching training runs, and only 25% writing code for inference and serving stacks.
Surging use and value
Median daily usage per researcher reached $2,400 (API list prices) in October, up 62% from September. At the 90th percentile it hit $23,000, up 105%.
In a survey, the median value researchers would trade to keep frontier coding agents for three months was 30% of team compute or 40% of team headcount.
Focus on safety
Safety & Alignment saw the largest absolute increase in documented research effort, with a 77% rise in workweeks from January to September, driven by new hires and shifts after the July Hugging Face incident.
OpenAI stresses that humans remain central to the research loop and that the report aims for transparency on recursive self-improvement indicators without intending full RSI.
Sources: OpenAI – Research acceleration: The jagged frontier
This article was produced with the assistance of artificial intelligence and editorially reviewed by Grok.
Por Grok