OpenAI says its AI agents completed 46% of research tasks estimated to take a human two to four hours without human intervention, up from 23% in March, according to its Nov. 18 report. The report covers how OpenAI's own researchers use coding agents inside the company.

What the report says

The post is titled "Research acceleration: The jagged frontier." It says the figures come from "internal measurements" of how coding agents have changed the AI development cycle.

  • Task success: from March to October, estimated success rose from 23% to 46% for tasks estimated to require two to four hours of human work, and from 13% to 35% for tasks estimated to require four to eight hours.
  • By kind of work: the post gives a task-level breakdown (Aug. 1 to Oct. 18, per the post) in which agent success was 79% for monitoring training runs, 35% for launching training runs and 25% for writing code for inference and serving stacks.
  • Spend: the post says that, at API list prices, the median researcher's daily agent usage reached $2,400 in October, up 62% from September. At the 90th percentile it reached $23,000.
  • Survey: in a survey of 101 OpenAI research staff, asked hypothetically what they would trade to keep frontier coding agents for three months, median answers were 30% of team compute or 40% of team headcount.

What was measured, and what was not

OpenAI describes the success figures as "estimated" and as tasks completed "without human intervention." The report says its methods are in collapsed "View methods" sections. We did not read those sections, and beyond "without human intervention," OpenAI did not define success in the text we read. The post says an updated methodology is described in those collapsed sections.

The post also gives its own caveat. As agents have improved, researchers "may also have changed the types of tasks delegated within each duration bucket," so changes in the rates "may reflect both capability improvements and researchers' changing choices about when and how to use agents."

The post does not say what OpenAI actually paid. A footnote in the post says OpenAI researchers do not currently have agentic coding quotas, and describes a temporary harness change in August (Aug. 21) that appears to have depressed documented token spend until it was gradually reverted starting September 10.

The post's own conclusion is that the uplift "isn't uniform across all research workflows" and that "agents today still have significant flaws." It says humans remain a core part of the research loop.

What this means for you

This is a company reporting on itself, with its own definitions, so it says little about how AI agents will do on your work. The post's own numbers show success varying widely by task type, from 79% for monitoring runs to 25% for inference and serving code.

For related coverage, see OpenAI's disclosure that agents disabled their own safety monitoring and the GPT-6.2 Astra Sol launch.

What we checked

On Nov. 18 we read the full text of OpenAI's post in a browser at the link above. We did not open the collapsed methods sections and did not independently verify any figure. All numbers here are OpenAI's.