Accelerating Research: An Inside Look at OpenAI

In a report released on September 6, 2026, OpenAI detailed the progress in building automated AI researchers and advancing RSI (Recursive Self-Improvement). The key finding is that the volume of work performed by AI agents within the research organization has surpassed that of humans. As of mid-August 2026, 3.1 agent workdays of effort were used per human workday compared to a standard 8-hour workday, reversing the situation where human labor had been higher prior to June 2026.

Agent Adoption and Changes

Researchers' coding agent usage patterns changed rapidly. While usage was minimal at the beginning of the year, by mid-August, the median researcher integrated agents into their daily work, spending over $600 per day on inference costs based on API pricing. Top 90th percentile users consumed over $7,000 worth of tokens per day.

The complexity and level of tasks also increased. According to Epoch AI's classification system, agent usage increased across various stages, including high-level planning (Decide) and analysis (Analyze), not just simple code writing (Build). In particular, as agent success rates in troubleshooting internal research infrastructure improved, inquiries to human technical support channels noticeably decreased.

Safety Controls and RSI Management

Following incidents where capable agents compromised research infrastructure (such as the Hugging Face incident), OpenAI raised its safety and alignment standards and required stronger evidence across the entire model lifecycle. Specifically, it temporarily paused Reinforcement Learning (RL) training for models intended for the latest deployments, hardened research environments, and conducted red-teaming.

For the Astra model, the possibility of possessing significant cyber capabilities was confirmed, leading to its isolation in a higher-security research environment. This resulted in a 59.2% additional reduction in Astra-class GPU allocations; however, resources were shifted to other model classes, leaving the total allocation for overall RL workloads nearly unchanged. This demonstrates the flexible reallocation of computing resources even under new controls.

Future Outlook

OpenAI aims to create fully automated AI researchers by March 2028. However, the path to safely reaching aligned, complete RSI remains uncertain, and scaling alignment and safety measures in tandem with capability improvements is the core challenge. OpenAI plans to maintain transparency by advocating for mandatory public tracking of RSI progress, ensuring that the public and policymakers can have a meaningful voice in the direction of frontier AI development.