OpenAI's AI agents complete 46% of multi-hour research tasks autonomously

Covered by 2 sources
Read onTechRadar· 3 min
OpenAI announces its AI agents now finish 46% of research tasks lasting two to four hours without human help, doubling their performance since March 2026. However, the agents face challenges with core software engineering tasks such as writing inference code. Usage of these agents per researcher has increased significantly, with success rates varying by task complexity and duration.

What happened

OpenAI announced that its AI agents can now autonomously complete 46% of research tasks estimated to require two to four hours, up from 23% in March 2026. For tasks expected to take four to eight hours, the agents' autonomous completion rate rose to 35%. The agents perform best in communication and analytical assignments, but face difficulties with core software engineering tasks, such as writing inference code. Researchers have increased their use of these agents, with spending and usage per researcher rising sharply.

Key details

  • AI agents completed 79% of training-run monitoring tasks autonomously.
  • Success rate for writing inference and serving code was 25%.
  • Communication tasks saw a 69% autonomous completion rate; analytical assignments reached 58%.
  • Median daily spending on coding agents reached $2,400 in October 2026, a 62% increase from September.
  • Documented contributions to Safety and Alignment research rose 77% from January to September 2026.

Why it matters

Researchers using OpenAI's AI agents can now automate a larger share of multi-hour research tasks, reducing the time needed for certain assignments. However, technical work like software engineering still requires significant human oversight, and researchers must address persistent issues with reliability and judgment. Increased agent usage also means higher costs for research teams, and more focus on safety and alignment work as AI systems become more capable.

Also covered by 1 source

More on this story

  1. OpenAI's own researchers now say they value their AI tools more than nearly half their colleagues
    · The Decoder
  2. Research acceleration: The jagged frontier
    · Open AI News
#openai#research-automation

More in AI Research