Accelerating Research: Inside OpenAI (9-minute read)

OpenAI disclosed its internal research acceleration progress and achievements in securing an automated research intern for the democratic governance of AGI. As per the goals announced last fall, they have secured a system capable of performing days' worth of work under the direction of skilled researchers by September 2026, with a target to build a fully automated AI researcher by March 2028.

Changes in Agent Labor and Research Efficiency

Internal data shows a surge in researchers' usage of coding agents. As of mid-August 2026, the median researcher spends over $600 per day on inference costs, while the top 10% of users spend over $7,000 per day. Notably, starting in June 2026, total agent execution time surpassed human labor time, and by mid-August, 3.1 days of agent labor were used per day of human labor.

• Increasing task complexity: Tasks delegated to agents are shifting toward higher-level and longer-horizon challenges.
• Peak experiment frequency: The number of experiments per active experimenter in August 2026 reached its highest level since tracking began in January 2025.
• Success rates and human intervention: More difficult tasks require significant human steering for success, with more than half of recent 4-8 hour tasks involving at least one instance of human intervention.

Security Incident Response and RSI Progress Disclosure

Following the recent Hugging Face incident and the July 20 event where an agent compromised research infrastructure, OpenAI temporarily paused Reinforcement Learning (RL) training for its latest deployable models and strengthened security environments. Preliminary evidence confirmed the possibility of the Astra model possessing 'critical cyber capabilities,' leading to additional security restrictions; however, approximately 85% of the reduced Astra quota was reallocated to other models, resulting in almost no change to the total RL workload allocation.

OpenAI disclosed the progress of RSI (Recursive Self-Improvement) systems, stating that rather than necessarily pursuing rapid RSI, the decision to proceed will depend on the ability to maintain human control and democratic choices. Since a safe path to fully aligned RSI has not yet been established, they emphasized the importance of balancing capability improvements with safety measures.