Anthropic and OpenAI Race Toward Recursive AI Self-Improvement

AI research firms Anthropic and OpenAI are competing to achieve recursive self-improvement, a state where AI systems automate their own research and coding to accelerate development. Anthropic co-founder Jack Clark reported that the company's AI model, Claude, now writes 80% of its code, resulting in an eight-fold increase in code production per person. OpenAI has set a target to fully automate its AI researchers by March 2028 and recently released GPT-5.3 Codex, a model that contributed to its own development from start to finish.

Despite the competitive race, leadership at both firms have expressed concerns regarding an intelligence explosion and the potential loss of human control. OpenAI Chief Scientist Jakub Pachocki and Anthropic Chief Science Officer Jared Kaplan have both called for international coordination to slow the process. Kaplan has integrated these warnings into Anthropic's Responsible Scaling Policy.

Skeptics, including Gary Marcus and Arvind Narayanan, argue that these developments represent faster coding rather than true recursive improvement. They suggest that progress remains limited by physical bottlenecks, such as data availability and computing power.