Anthropic reports AI systems now accelerating their own development, raising recursive improvement concerns

Anthropic reports its AI systems are now accelerating their own development, a trend that could lead to recursive self-improvement. The company's engineers ship 8x more code than before, with over 80% of it authored by its AI, Claude. This acceleration, also seen in public benchmarks, raises concerns about humans losing control, prompting Anthropic to suggest a verifiable slowdown in frontier AI development to allow safety research to advance.

AI-accelerated AI development

• ▪AI models have advanced from scoring in the low single digits to saturating the SWE-bench software engineering test in two years
• ▪Anthropic reports its internal data shows AI systems are accelerating the development of more advanced AI
• ▪The length of tasks AI systems can reliably complete on their own has been doubling roughly every four months, an acceleration from a previous trend of doubling every seven months

Recursive self-improvement trajectory

• ▪In March 2024, Claude Opus 3 could complete 4-minute software tasks; by March 2026, Claude Opus 4.6 managed 12-hour tasks
• ▪If current trends hold, AI systems could be capable of tasks that take a person weeks to complete in 2027
• ▪The trend of AI accelerating its own development could lead to recursive self-improvement, where an AI system autonomously designs and develops its own successor

Claude coding productivity metrics

• ▪In a March 2026 poll of 130 employees, the median respondent estimated they produced around four times as much output with Mythos Preview than without AI models
• ▪Anthropic engineers now ship eight times as much code per quarter as they did from 2021 to 2025
• ▪As of May 2026, over 80% of the code merged into Anthropic’s codebase was authored by its AI model, Claude, up from low single digits before February 2025

Engineering automation at Anthropic

• ▪Proposed changes to Anthropic's codebase are now reviewed by an automated Claude reviewer that looks for bugs and security flaws
• ▪Claude's success rate on open-ended engineering tasks reached 76% in May 2026, an increase of 50 percentage points in six months
• ▪A retrospective analysis found an automated Claude review would have caught roughly one-third of the bugs behind past incidents on claude.ai

Research task execution capabilities

• ▪In an April 2026 demonstration, Claude-powered agents solved an open-ended AI safety problem, recovering 97% of a performance gap over 800 hours
• ▪In an experiment to optimize code, Claude Mythos Preview achieved a ~52x speedup by April 2026, while a skilled human researcher would need 4-8 hours to reach a 4x speedup
• ▪In the same AI safety demonstration, two human researchers recovered roughly 23% of the performance gap over about a week

AI control risks

• ▪Anthropic would slow down or pause its own development if other frontier AI developers also did so in a verifiable manner
• ▪Anthropic stated that a slowdown in frontier AI development would be beneficial, allowing societal structures and alignment research to catch up
• ▪Full recursive self-improvement might increase the risks of humans losing control over AI systems