Recently, Anthropic published an important article on its blog, stating that the development of artificial intelligence (AI) is gradually accelerating and even beginning to self-propel towards self-improvement. Self-improvement refers to AI systems that can autonomously design, train, and evaluate to continuously improve their next-generation versions. Once this development takes shape, it will completely change the speed at which AI progresses.
The article mentions that AI is now approaching, and even surpassing, human capabilities in execution. For example, according to public benchmark tests, the time AI takes to complete independent tasks doubles every four months: it is expected that by 2026, Claude 4.6 will be able to handle tasks for up to 12 hours. Additionally, internal data from Anthropic shows that by May 2026, over 80% of code contributions were completed by Claude, a figure that was just a single digit two years ago.
In terms of handling complex tasks, Claude's performance is also remarkable. A survey involving 130 people found that the median work output of participants assisted by Claude was four times higher than when no AI was used. In April 2026, Claude successfully fixed over 800 API errors, significantly reducing the rate of errors.
However, despite significant progress in execution capabilities, Anthropic still emphasizes that humans have clear advantages in research judgment, problem selection, and trust in results. While AI's capabilities in execution are gradually approaching those of humans, AI still struggles with deciding what to do and how to control the direction.
