OpenAI's success rate in coding intelligent agent research tasks has improved, but the cost of using APIs has increased significantly.
A research report released by OpenAI shows that the success rate of coded agents in research tasks expected to take two to four hours increased from 23% in March to 46% in October, while the success rate for tasks taking four to eight hours increased from 13% to 35%. Humans remain at the heart of the research cycle. Based on API pricing, the median daily usage by researchers reached $2,400 in October, a 62% increase from September; the 90th percentile reached $23,000, a 105% increase. Between January and September, the number of working weeks for safety and alignment tracking increased by 77%, the largest absolute increase among all major research areas. OpenAI advocates that labs should track and publicly disclose Recursive Self-Improvement (RSI) metrics.