OpenAI has announced the launch of a recursive self-improvement loop, with AI now beginning to participate in AI R&D. The cost for internal researchers using AI agents is rising rapidly—heavy users are spending up to $23,000 per day—and the corresponding agent workdays per person are growing substantially. Researchers are even willing to cut some staff and reduce core compute to retain frontier coding agents. In the latest R&D efforts, AI can already propose key scientific insights and demonstrates a preliminary sense of scientific taste, accelerating the R&D feedback loop. Current AI capabilities remain severely uneven, with zero-intervention success rates on complex R&D tasks still limited. Humans continue to play the core role of review and gatekeeping, but their advantageous position cannot be guaranteed in the long term. OpenAI calls for mandatory disclosure of core recursive self-improvement metrics. Currently, only 2% of R&D compute is used to deploy internal agents; model capabilities are improving at high speed, and technological breakthroughs may arrive at any time.
