AI labs turn self-improvement into a 2026 benchmark

Recursive Self-Improvement, long rooted in I. J. Good’s 1965 idea of an “intelligence explosion”, has shifted from theory into measurable lab programs in 2026. ICLR hosted a dedicated Workshop on RSI in April, while Recursive Superintelligence emerged in May with $650 million in financing to pursue models that can identify weaknesses and redesign themselves.

Anthropic reported that as of May 2026, more than 80% of code merged into its codebase was written by Claude, up from single digits before Claude Code launched in February 2025. In the second quarter of 2026, code merged per engineer per day was 8 times that of 2024. Claude also improved sharply on optimization tasks, moving from about 3 times speedup in May 2025 to about 52 times by April 2026.

OpenAI said on September 6 that it had reached its goal of an intern-level research assistant, with researchers using about 3.1 corresponding “agent workdays” for every 1 human workday. Domestic efforts are emphasizing engineering loops: MiniMax said M2.7 ran more than 100 rounds of scaffold improvement, while Zhipu said GLM-5.3 built GLM-5.3-Flash inference service on more than 100,000 domestically produced chips and lifted throughput to 3 times the initial baseline in less than two weeks.