Artificial intelligence is increasingly being used to help build the next generation of AI systems, with Anthropic reporting a sharp rise in the role played by its Claude model in the company’s own research and development.
According to Anthropic’s latest measurements, Claude was at the “leads” level for 26% of the company’s measured AI R&D work in August 2026. At this level, AI can complete most of a task from a high-level prompt while a human continues to supervise the work.
The figure was below 1% in February 2026. More than 90% of Anthropic’s measured AI R&D work now involves AI at the “collaborates” level or above. However, Anthropic said Claude has not reached full autonomy in any measured area of AI R&D.
How Anthropic Measures AI-Led R&D
Anthropic has developed an R&D Automation Index to measure how much of its AI research and development work is being performed with the help of AI.
The system uses an automation scale developed by Epoch AI, ranging from AL0, where AI is not involved, to AL5, where an AI system performs the work fully autonomously without a human in the loop.
At AL3, AI “collaborates”, meaning it can carry out substantial parts of a task under close human direction. At AL4, AI “leads” the task and can complete most of it from a high-level instruction while a human supervises.
Anthropic reported that Claude reached AL4 for 26% of its measured AI R&D work by August. The company stressed that this does not mean Claude is independently building its own successor. Human oversight remains part of the process, and no measured category had reached AL5.
Why AI Developing AI Matters
Anthropic is using these measurements to track progress toward what is known as recursive self-improvement, in which AI systems could eventually contribute to developing increasingly capable successors.
The company said that stage has not yet been reached. However, the growing use of AI for research, coding, testing and engineering could change how quickly new AI systems are developed.
Anthropic argues that measuring the amount of AI involvement in AI R&D can therefore provide a useful way to track changes in the development process. The company has also called for other frontier AI developers to publish similar measurements regularly so that progress can be followed over time.
Around 30,000 AI Agents Running Inside Anthropic
Anthropic also disclosed figures on the use and monitoring of AI agents across its research and engineering operations.
In August 2026, approximately 30,000 agents were carrying out research and engineering work simultaneously on the company’s most-used internal platform. Anthropic said the measurements cover that platform only.
Every action taken by these agents passes through an online monitoring system before execution. An offline monitoring system also reviews their activity after the action has taken place.
Anthropic analysed more than 1 billion decisions made by research and engineering agents during August. About 0.002% of those decisions, or roughly one in 47,000, were blocked by the online monitoring system.
The offline system flags around one to two transcripts per 1,000 for additional review. Anthropic said roughly 50 of the highest-priority cases are escalated to human reviewers each week.
The company says these measurements are intended to show how oversight systems need to develop as AI agents take on longer and more consequential research and engineering tasks.
How Much Computing Power Is Used for AI Safety?
Anthropic also published figures on the share of computing resources being used for safety-related research.
During a one-week measurement period in July 2026, about 6% of the computing resources used for AI R&D went toward safety work. When looking specifically at compute associated with AI-driven AI R&D, the figure was approximately 12%.
Anthropic described these figures as conservative estimates. It also noted that computing resources do not provide a complete measure of safety investment because some safety research depends more heavily on human expertise than on computing power.
Anthropic Calls for More Transparency From AI Labs
Anthropic said frontier AI companies should consider publishing comparable data on a regular basis to provide a clearer picture of how quickly AI development is changing.
The company acknowledged that comparisons between laboratories would be difficult because there is currently no standardised methodology for measuring AI-led R&D.
Another challenge is that AI systems are being used to evaluate AI-generated work. Anthropic noted that a model acting as a judge could make some of the same types of errors as the system being evaluated.
The company said independent third-party verification could help address these concerns. It also said other AI developers could publish measures covering the extent of AI involvement, agent monitoring and the resources devoted to safety work.
Anthropic’s latest figures show that AI is already playing a growing role inside the process of developing new AI systems. At the same time, the company’s own measurements indicate that the measured work has not reached full autonomy, with human oversight still involved across the process.