OpenAI opens its internal agent data: autonomous success on two-to-four-hour research tasks rose from 23% to 46%
Synopsis
OpenAI published internal measurements of how coding agents are used in its own AI research: from March to October 2026, estimated success without human intervention rose from 23% to 46% for tasks estimated to need two to four hours of human work, and from 13% to 35% for four-to-eight-hour tasks. The uplift is uneven, reaching 79% for monitoring training runs but only 25% for writing code for inference and serving stacks, and humans remain in the research loop. At API list prices, the median researcher's daily agent spend reached $2,400 in October, up 62% from September.
Interpretation
OpenAI published a set of internal measurements describing how coding agents are changing its own AI development cycle, organized around inputs, process and outputs. The earlier September report gave an earlier version of the analysis; this update adds quantitative and qualitative measures of AI uplift across research domains and of how humans remain involved. The report describes the picture as broader but still partial, and states that its historical trends and metrics involve significant human-AI collaboration and may not predict future AI-driven progress.
The share of research tasks agents complete on their own rose markedly over a few months, but the gap between task types is large. Binned by estimated human hours, zero-intervention success rose from 23% in March to 46% in October for two-to-four-hour tasks, and from 13% to 35% for four-to-eight-hour tasks. Under Epoch AI's O*NET AI R&D taxonomy, from August 1 to October 18, 2026, estimated success was 69% for Communicate, 58% for Analyze and 43% for Build; at task level, 79% for monitoring training runs and 72% for answering technical questions, versus 25% for writing code for inference and serving stacks and 29% for building evaluation, reinforcement-learning and agent environments.
Researchers' intensity of agent use rose quickly over the same period, and heavy users pulled further ahead of the median. At API list prices, the median researcher's daily agent spend reached $2,400 in October, up 62% from $1,500 in September; the 90th percentile reached $23,000, up 105% from $11,000. Agent workdays per researcher rose from 0.5 to 5.7, about elevenfold; pretraining and posttraining teams averaged roughly eight agent workdays per researcher per calendar day and scaling and infrastructure about four; by mid-October more than two-thirds of included researchers averaged a daily peak of at least four concurrent agents and more than 10% at least sixteen.
Researchers shifted substantial time toward Safety and Alignment, the largest absolute increase among OpenAI's major research areas. Measured in documented workweeks classified from company-public Slack posts, pull requests and internal documents, Safety and Alignment grew 77% from January to September, nearly doubling. The report attributes the growth to new hires and existing researchers shifting focus, with growth picking up significantly after the response to the July 2026 Hugging Face incident.
Perspective
The measurements cover coding-agent use inside OpenAI's research organization during 2026, with task success rates binned by estimated human hours and classified under Epoch AI's O*NET AI R&D taxonomy, and spend converted at API list prices. The report addresses readers concerned with AI R&D acceleration and governance, and describes itself as offering a broader but still partial picture.
The report notes that as agents improve, researchers may change which tasks they delegate within each duration bucket, so success-rate changes may reflect both capability gains and changing usage choices; the survey is self-reported with an ad-hoc sampling strategy; the boundaries between research areas, especially posttraining versus safety and alignment, are nontrivial and the definitions may be iterated; and the historical trends and metrics involve significant human-AI collaboration and may not predict future AI-driven progress.
