Tibo Sottiaux, head of OpenAI Codex, recently stated that the company has optimized its underlying inference architecture through powerful models, reducing the operating cost of its Luna model by approximately 80%. He explicitly characterized this process as an early form of Recursive Self-Improvement.