Claude AI

The company’s research arm says Claude authors more than 80% of its production code and warns that AI designed to build its own successors could arrive before governments can respond.

 

The Anthropic Institute published an analysis this week showing that AI is already accelerating its own development and called on governments, rival labs, and civil society to build the systems needed for a global AI pause. Anthropic said it would stop if other leading labs agree to pause development, as long as there is a way to verify everyone adheres to the agreement.

The report uses publicly available benchmarks alongside previously undisclosed internal Anthropic data.

The central finding: AI is no longer just a tool used in development. It is now a driver of it.

 

AI already does most of the work at Anthropic

Claude now writes more than 80% of the code that goes into Anthropic’s systems. Engineers are now reviewers rather than writers. That shift increased the amount of code each engineer ships per day by 8x since 2024, and employees say they get around 4x as much work done.

The range of tasks AI can handle without human help doubles every four months. In March 2024, Claude could complete tasks a human finishes in four minutes. By April 2026, it was handling work that takes a skilled person 12 hours.

On open-ended research tasks, Claude’s success rate reached 76% in May 2026, up 50 percentage points from six months earlier. On deciding what to research next, Claude now outperforms human researchers 64% of the time in live sessions, up from 51% in November 2025.

 

Once AI builds itself, controls become much harder to add

The analysis lays out three possible futures. In the first, AI capability growth stalls and today’s AI spreads gradually through the economy. In the second, AI handles more and more of the work while humans still set the direction. In the third, AI systems design and train their own successors without meaningful human input. Researchers call this recursive self-improvement. Anthropic says the first scenario is unlikely.

If AI systems can design their own successors, the pace of progress is no longer limited by how fast humans can conduct research. It becomes limited only by computing power. At that point, the controls needed to keep AI systems safe become much harder to build after the fact.

 

A pause only works if everyone can prove they stopped

Anthropic said that an individual company can easily pause its development, but that would only change who leads the race, not alter the pace of development. A real coordinated pause requires a way to verify that all frontier AI developers stop at the same time. The report points out that AI training runs are far easier to hide than missile silos, and that comparable arms control treaties took decades to build.

Anthropic says it will organize conversations with policymakers, researchers, civil society, and other AI companies in the coming months and will publish the results.