AI helps AI: Claude now leads 26% of Anthropic research
AI is starting to build AI. Anthropic says Claude now leads 26% of its R&D work, up from less than 1% in February. Here's all that you need to know!
AI is starting to work on AI. The technology that once helped researchers write code, analyse data or answer questions is increasingly being used to do the same work involved in building the next generation of AI models.
Anthropic's latest figures offer a glimpse of what that looks like inside a frontier AI lab. The company says Claude now leads 26% of its AI research and development work, up from less than 1% in February.
So, what does it actually mean when an AI model helps build better AI?
AI moves from assistant to researcher
Building an AI model involves far more than training it on data. Researchers need to write and test code, analyse results, evaluate model behaviour, identify problems and run experiments. AI can now take on parts of this workflow.
Anthropic measures this using the Anthropic R&D Automation Index, based on an automation scale developed by Epoch AI. The scale runs from AL0, where AI has no involvement, to AL5, where an AI system can complete work fully autonomously without a human in the loop.
Claude's 26% figure represents work at AL4. At this level, the AI can complete most of a task from a high-level instruction, while a human continues to supervise it. That distinction matters.
Claude is not independently running Anthropic's research programme, and the company says it is not yet fully autonomous in any measured area of its AI R&D. Still, more than 90% of Anthropic's AI R&D work is now at a level where AI either teams with humans or leads the task.
How can AI improve another AI?
Imagine a researcher developing a new model. Instead of manually writing every experiment, they could ask an AI agent to create code, run tests, compare results and suggest changes. The researcher can then review the work and decide what happens next.
The process can be repeated across thousands of experiments. This matters because AI research contains many tasks that are well suited to automation. Faster coding, testing and evaluation could allow researchers to explore more ideas and identify problems sooner.
It also creates the possibility of a feedback loop. Better AI can help researchers build better models, which could then become more capable at helping with AI research.
This is where the idea of recursive self-improvement enters the conversation. It describes a future scenario in which an AI system could help develop a significantly more capable successor with decreasing human involvement.
Anthropic's current numbers do not show that this has happened. Instead, they show increasing automation within the development process.
The rise of AI research agents
This work is also increasingly being handled by AI agents rather than simple chatbots. Anthropic said around 30,000 AI agents were carrying out research and engineering work on its most-used internal platform at any given time in August 2026.
These agents can take actions across tasks instead of simply responding to prompts. Anthropic said their actions are checked by an online monitor before execution, usually within seconds, followed by additional offline monitoring.
In August, the company analysed more than 1 billion agent decisions and said 0.002% were blocked by its online monitor.
The new challenge is oversight
As AI takes on more of the work involved in developing AI, the question is no longer only how quickly models can improve.
It is also how well humans can monitor the process. Anthropic said about 6% of computing power used for AI R&D went towards safety work during a sample week in July, compared with about 12% for AI-driven AI R&D. The company also cautioned that computing power is an imperfect measure of safety effort.
The numbers point to an important change in AI development. Models are no longer just the products of research. They are increasingly becoming part of the machinery used to produce the next models.
The more that process is automated, the more important it becomes to know what the AI is doing, where humans remain in control and whether the systems being built can be reliably evaluated.