What matters in AI.

Subscribe

OpenAI agents do 46% of research tasks of 2 to 4 hours

The rate was 23% in March 2026, but the only result is an internal test of OpenAI.

OpenAI gives new results for its AI agents. From March to October 2026, the agents did 46% of research tasks of 2 to 4 hours with no help from a person. The rate was 23% in March. OpenAI tells that its researchers continue to give the direction and the judgment. The numbers come from an internal test of the company.

What is new

For tasks of 4 to 8 hours, the rate went from 13% to 35%. Researchers use coding agents more. The median spend for each day was $2,400 in October.

The numbers

  • Agents did 79% when they monitor training runs, which make a model.
  • Agents did only 25% when they write software for inference and serving systems.
  • Communication tasks gave 69% and analytical tasks gave 58%.

What is not known

Work that looks done can hold mistakes, and a person must correct it. OpenAI tells that its numbers do not show faster science. Its index of model results uses different tests from the index of Epoch AI.

Insight: Test results differ with the person who does the test

OpenAI reports that its agents do 46% of tasks of 2 to 4 hours. Other tests give lower results: the top coding agent does only 26.4% of tasks fully, and the top agent does 46% of clinical tasks. This reads as a difference: a company that tests the agents that it made shows more than external researchers do. Tell the company that made your agent to show tests on your hard tasks.

Sources

  1. AI agents ace 79% of training-run checks but stumble on core software engineeringTechRadar
AI MATTER · NEWS · AI MATTER · NEWS ·21 NOV2026

Posted

Tags

Learn the terms in this story

More in Agents

All Agents news