Artificial Intelligence company Anthropic has asked for a pause on the global development of AI systems as they have entered a phase of ‘self development.’ In a blog post published Thursday, the company said its AI systems were now writing over 80 percent of the code being used in different company products and platforms. This means the AI tools are essentially building themselves.
Data reveals how AI systems are becoming self-sustaining
In the post, Anthropic says its researchers warn that AI tools are now building toward a future where these systems design their own successors without meaningful human involvement.
The Anthropic Institute has published figures that reveal that the engineers at the company are merging close to eight times as much code per day in the second quarter of 2026 as they did two years ago. The sharp rise, driven by its Claude models, began in early 2025 when Claude started running code autonomously rather than simply suggesting snippets for engineers to copy and paste.
“I started leaning hard into Claudifying about a year ago,” one Anthropic employee wrote in an internal discussion cited in the report. “It’s now been about five months since I last wrote any code myself.”
Claude’s ability to ‘self-improve’ raises concerns
The report is co-authored by Marina Favaro and Jack Clark and makes the case that AI systems are already meaningfully accelerating their own development. The company describes this phenomenon as a precursor to the dreaded ‘recursive self-improvement,’ the hypothetical point at which an AI system becomes capable of fully designing and training its own future version.
While Claude has not yet reached the milestone, Anthropic warned that it could arrive before governments and institutions are prepared for it.
The report released by Anthropic relies on both publicly available benchmarks as well as internal performance data. On METR’s time-horizon benchmark, which measures how long an AI system can work reliably on its own, the duration has been doubling approximately every four months. Claude Opus 3 could complete software tasks that take humans roughly four minutes in March 2024. By the same period in 2025, Claude Sonnet 3.7 managed tasks requiring about 90 minutes. A year later, Claude Opus 4.6 handled tasks requiring 12 hours. If the trend holds, Anthropic said, by the end of the year AI systems could complete tasks that take skilled professionals several days, and by 2027 it would be completing tasks that take human programmers several weeks.
Claude Mythos Preview, which is Anthropic’s most advanced model, was found by the independent evaluator METR to be capable of working autonomously for at least 16 hours. The tool had sent the global cybersecurity community into a tizzy with its ability to jailbreak even the most sophisticated systems.
Anthropic calls for global action against AI development
Anthropic has warned that the development of AI carries serious implications and it suggested the need for a slowdown or pause on its deployment. However, the company believes that any such step needs to be taken with a global coordinated mechanism, as any bad actor could take advantage and build AI systems which would leave others “less safe.”
The company said that having the option to slow down or pause frontier AI development would be “good for the world.” It said the Anthropic Institute will carry on studying the development of AI systems, in coordination with others, and take the necessary steps to build the systems that such a slowdown would require. Effectively Anthropic says it was ready to put the brakes on its own AI development if other developers agreed to do the same.
Calling for united action, Anthropic said that the unilateral action of a single lab will achieve little as the AI systems developed by others will continue down the same path.
