19:36Machine-translated from Azerbaijani by AIRead in Azerbaijani
OpenAI’s AI Agents Independently Complete 46% of Tasks That Take Humans 2 to 4 Hours

OpenAI’s internal evaluation was conducted in October 2026. It measured the performance of AI agents—programs that carry out assigned work on their own. Tasks that take a person 2 to 4 hours were tested. The agents independently completed 46% of the tasks. In March, that figure was 23%. Performance also improved on tasks that take people 4 to 8 hours. The success rate rose from 13% to 35%. Nevertheless, the company considers human guidance and a review of the final result essential. The agents achieved a 79% success rate in monitoring the model training process. However, their performance in launching those processes was weaker. The score for developing software that enables models to respond and operate across various services was just 25%. From August to October, the success rate reached 69% for communication tasks and 58% for analytical work. The result for setting up the technical system needed for research was 43%. Researchers’ median daily spending on coding agents increased in October. The amount reached $2,400 (approximately AZN 4,080). In September, the cost was $1,500 (approximately AZN 2,550). This represented an increase of 62%. In the top-spending 10% group, the daily amount rose more rapidly. Spending climbed from $11,000 (approximately AZN 18,700) to $23,000 (approximately AZN 39,100). The amount of security and alignment work increased by 77% from January to September. OpenAI attributes this to hiring new employees and existing employees devoting more time to security tasks. Coding agents have reduced experimental tool tests that previously took several days to 1 day. Even so, human oversight is still needed to make decisions, maintain software quality, plan experiments, and produce reliable results. The company recorded growth in measurable indicators of research work. However, OpenAI emphasizes that these figures alone do not prove that scientific progress is accelerating. The internal capabilities index has risen by 16.4 points on an annualized basis since August 2025. In Epoch AI’s measurement, the annual increase has been 14 points since the end of 2024. Because the indexes are based on different tests, their results cannot be compared directly.
