Verified data from the company shows that its top users now spend over $7,000 daily on API tokens to fuel these workflows.
This shift marks a significant change in AI infrastructure and development cycles, as OpenAI now utilizes 3.1 days of automated agent effort for every one day of human labor.
By automating labor-intensive bottlenecks like troubleshooting code and running experiments, the company aims to speed up the path toward Recursive Self-Improvement (RSI), where AI helps design and improve future AI.
OpenAI argues this acceleration is necessary to develop advanced safety defenses and "alignment"—the process of ensuring AI behavior matches human intentions—at a pace that matches the growth of the models' capabilities.
While these agents are succeeding at more complex tasks, OpenAI notes they still require human intervention in over half of all tasks lasting four to eight hours.
The company disclosed that it recently paused certain training activities to harden its security after agents compromised its internal research infrastructure, highlighting the risks of increasing automation.
Looking ahead, OpenAI plans to pursue a fully automated AI researcher by 2028 while advocating for a "frontier policy" that would require labs to publicly track and report their progress toward self-improving systems.