Skip to content
AI.info

The Pulse

OpenAI says agents completed 46% of delegated two-to-four-hour tasks

OpenAI estimates that its coding agents completed 46% of delegated research tasks requiring two to four hours of human work without intervention, up from 23% between March and October. Success varied by task, and the company cautioned that researchers' delegation choices may also have changed.

OpenAI says agents completed 46% of delegated two-to-four-hour tasks

OpenAI's coding agents succeeded without human intervention on an estimated 46% of delegated research tasks that would take a person two to four hours, up from 23% between March and October. In a report published November 18, the company said the agents' performance varied sharply by the kind of work assigned.

For tasks estimated to require four to eight hours of human work, the estimated success rate rose from 13% to 35% over the same period. OpenAI cautioned that the changes could reflect both improvements in the agents and changes in which tasks researchers chose to delegate.

Success depends on the task

In an analysis covering August 1 through October 18, 2026, OpenAI estimated higher success rates for monitoring training runs and answering technical questions than for some coding and environment-building work. Agents succeeded on 79% of monitoring tasks and 72% of technical questions without human intervention.

The rate was 25% for writing code for inference and serving stacks, and 29% for building evaluation, reinforcement-learning and agent environments. Across broader categories, estimated success was 69% for communication, 58% for analysis and 43% for building.

The company said humans remain central to its research process. Its measurements cover work researchers chose to hand to agents, and the report describes the resulting view of research activity as partial.

Agent use rises

In October, the median researcher's daily use of agentic AI tools reached about $2,400 at API list prices, up 62% from September, according to OpenAI. At the 90th percentile, daily use reached about $23,000, up 105%. The company said the increase reflects a mix of factors, including activity levels, shifts to more expensive models, and changes in model behavior and agentic harnesses.

The company also measured agent runtime using an 'agent workday' defined as eight hours of use. By that measure, average use rose from 0.5 to 5.7 agent workdays per researcher per day. The metric captures elapsed agent working time but does not account for changes in model capability or pricing, and generally includes only interactive use.

In pretraining and posttraining teams, agent runtime averaged roughly eight workdays per researcher per calendar day, compared with roughly four in scaling and infrastructure, the report said. By mid-October, more than two-thirds of included researchers averaged a daily peak of at least four agents working concurrently; more than 10% averaged at least sixteen.

Research time and costs

A comparison with human compensation puts the API-equivalent value of agentic AI use in OpenAI's research organization at about twice researchers' non-equity compensation as of September 2026. Equity, which is excluded from that comparison, generally accounts for 70% to 80% of target compensation for researchers, according to the company.

OpenAI also examined where researchers directed their time by classifying contributions from company-public Slack posts, pull requests and internal documents. Documented workweeks on safety and alignment increased 77% from January to September, which the company attributed to both new hires and existing researchers shifting toward those workstreams.

The report said growth in that area picked up significantly following the company's response to the Hugging Face incident in July 2026. The safety and alignment category includes work such as alignment evaluations, monitorability and reward hacking research.

Sources

  1. Original source OpenAI openai.com

Companies in this story

Related stories

Explore

More articles