On September 14, 2026, OpenAI announced that its best model optimized kernels and cut costs 80% on Luna and Terra. Recursive's roadmap starts with AI for AI research rather than physical sciences, targeting cheaper training and inference including local inference. A month later OpenAI reported an 18% improvement via recursive self-improvement, with GPT-5.6 improving itself. What remains open is how far that approach reaches beyond cost savings.