OpenAI said its internal AI agents now complete 46% of researchers’ tasks that take two to four hours without human intervention. The rate was 23% in March 2026.

For tasks lasting four to eight hours, the agents’ success rate rose from 13% to 35%.

At API list prices, median daily agent usage per researcher was valued at approximately $2,400 in October. That was 62% higher than in September.

Agents succeeded at 79% of tasks involving monitoring training runs. Their success rate for writing serving code was 25%.

The figures suggest AI is becoming a key tool in accelerating OpenAI’s development.