ACN · Presented by Levr
ACN WIRE

Agents

OpenAI: agents finish 46% of multi-hour tasks

OpenAI's second report on agents in its research

Written with AI assistance, reviewed by the ACN desk.

“OpenAI cautions in its conclusion that the data gathered so far from human-AI collaboration may not extrapolate linearly.”
The Decoder, The Decoder
0:00 / 0:00

The quick hit

  • OpenAI has published new data on how AI agents speed up its own research.
  • Agents working alone succeeded on an estimated 46% of two- to four-hour tasks.
  • That success rate was 23% in March, according to The Decoder.

OpenAI has published a second internal report on how AI agents speed up its own research, according to The Decoder.

The Decoder reports that agents working without human help succeeded on an estimated 46% of two- to four-hour tasks. That rate rose from 23% in March. Success was lower on complex engineering, such as inference and serving code at 25%. At API list prices, the median researcher would have used $2,400 of tokens per day in October.

OpenAI says it wants to track indicators of recursive self-improvement, where AI speeds up building AI. In an internal survey of 101 employees, the median researcher would rather give up 40% of team staff than lose coding agents for three months, The Decoder reports.

The Decoder lists several caveats. The survey is internal and not representative. The $2,400 figure is a theoretical API-price value, not money paid. OpenAI cautions that the data gathered so far may not extrapolate linearly.

“OpenAI cautions in its conclusion that the data gathered so far from human-AI collaboration may not extrapolate linearly.”
The Decoder, The Decoder

Sponsor

Sources

  1. The Decoder · Nov 19, 2026

Spot an error? Email corrections@aicodingnews.si. Fixes are logged on the corrections page.

From our sponsors

Sponsor

Sponsor

Also in

Sponsor

Sponsor