Recursive self-improvement blog post from June

Anthropic's June blog post presents internal and external data showing AI is already meaningfully accelerating its own development. Engineers at Anthropic now merge 8x more code per day than in 2024, with over 80% of merged code authored by Claude as of May 2026. Externally, AI task-completion horizons are doubling roughly every four months, with models progressing from handling 4-minute tasks in early 2024 to 12-hour tasks by mid-2026. Anthropic defines recursive self-improvement as an AI system fully autonomously designing its own successor and states this is not yet achieved but may arrive sooner than institutions are prepared for. The post frames this trajectory as carrying both significant upside in science and healthcare and meaningful risk of humans losing control over AI systems.