Claude Is Now Helping Build the AI That Comes After It — Anthropic Reveals a Striking New Milestone

The AI Building the Next AI

There is a new number coming out of Anthropic that is hard to ignore: 26%. The company says Claude now “leads” roughly 26% of its artificial intelligence research and development work. That does not mean Claude is sitting in an office somewhere independently deciding what Anthropic should build next. Humans are still involved, and Anthropic says Claude is not fully autonomous in any of the work measured. But the shift is remarkable because the number was just 1% in March. In only a few months, AI has gone from being a relatively small assistant inside the development process to taking the lead on a significant portion of the work. citeturn0search17turn0news5

That distinction matters. When people hear that “AI is building AI,” it can sound like a science-fiction scenario in which machines have completely taken over their own development. That is not what Anthropic is saying has happened. The company describes Claude as leading tasks while humans provide high-level direction and supervision. Still, the underlying trend is exactly what has made researchers increasingly interested in the idea of recursive self-improvement — a future in which AI systems become capable enough to meaningfully accelerate the development of their successors.

From Writing Code to Helping Decide What Gets Built

Anthropic has already been seeing Claude become deeply integrated into its engineering process. According to the company's own research, more than 80% of the code merged into Anthropic's codebase was authored by Claude as of May 2026. Anthropic also says the typical engineer was merging about eight times as much code per day in the second quarter of 2026 compared with 2024, with Claude doing much of the actual coding while engineers direct and review the work. citeturn0search18

The newer 26% figure takes the story a step further. Coding is one thing; helping with broader research and development is another. Anthropic says Claude's role can include work that requires longer chains of reasoning and more independent execution, although people remain responsible for supervision. The company says AI collaborated with humans on more than 90% of its research work in August, showing that the reality today is less “AI versus humans” and more a rapidly changing partnership between the two. citeturn0search17

Anthropic Has 30,000 AI Agents Working Internally

Perhaps the most eye-opening number is not even the 26%. Anthropic says around 30,000 AI agents were carrying out research and engineering work on its main internal platform at any given time in August. Every action was screened before execution, and the company reported that only about one in every 47,000 decisions was blocked during a month involving more than a billion decisions. citeturn0search17

Think about what that means for the modern AI lab. A few years ago, an AI model was primarily something researchers trained and users interacted with. Now, thousands of AI agents can work alongside the people developing the next generation of models, writing code, investigating problems and performing other technical tasks. The computers are no longer simply running the AI — increasingly, the AI is helping determine what those computers should do next.

And This Is Where the Safety Debate Gets Real

The timing of Anthropic's disclosure is particularly interesting. CEO Dario Amodei has recently argued that frontier AI development should be slowed enough for safety measures to catch up with capability improvements. His proposal came amid growing concern about increasingly autonomous AI systems and their potential misuse. citeturn0news22turn0news1

Anthropic's latest numbers give that debate something more concrete to talk about. If AI systems are increasingly capable of contributing to the development of their successors, then the question is no longer simply how powerful the next model will be. Researchers also have to ask how quickly AI development itself can accelerate once models become productive participants in the process.

Anthropic says it is tracking these numbers precisely because of that concern. The company wants to measure how close AI systems are getting to the point where they could substantially accelerate their own development, while giving the public a clearer picture of what is actually happening inside frontier AI companies.

Humans Are Still in the Loop — For Now

It is important not to exaggerate the announcement. Claude has not independently taken over Anthropic's AI research program, and the 26% figure does not mean that 26% of Anthropic's company is being run by machines. The measurement is based on specific research and development tasks, and humans remain involved in setting goals, supervising agents and reviewing their work. citeturn0search17

But the direction is difficult to miss. In March, Claude led about 1% of the measured R&D work. By August, Anthropic says that figure had reached 26%. At the same time, Claude is writing most of the code merged into Anthropic's own codebase, while tens of thousands of AI agents are being used internally.

The interesting question is therefore not whether AI has suddenly become completely independent. It hasn't. The more important question is how long it will take before the human role changes from doing the work to primarily directing, checking and deciding what the AI should do.

The AI Race Is Starting to Look Different

This could become one of the defining changes in the AI race. OpenAI, Google, Anthropic and other frontier labs are not simply competing to build smarter chatbots anymore. They are building systems that can increasingly perform the research, coding and experimentation required to create the next systems.

That creates a potentially powerful feedback loop: better models can help engineers work faster; faster engineering can produce better models; those better models can then contribute even more to the next development cycle.

Nobody knows exactly where that loop ultimately leads. Anthropic itself is careful not to claim that recursive self-improvement has already arrived. But its latest numbers show that the boundary between “AI assistant” and “AI developer” is becoming increasingly difficult to draw.

And perhaps that is the real story behind the 26% figure. The most important AI being built today may not simply be the AI that answers our questions. It could be the AI sitting behind the scenes, helping researchers build whatever comes next.