Anthropic Says Claude Now Handles 26% of Its AI Research and Development

Anthropic says its Claude artificial intelligence models are now playing a much larger role in the company's own effort to develop future AI systems.

As of August 2026, Claude was able to lead 26% of Anthropic's artificial intelligence research and development work, according to a new measurement framework published by the company. That figure represents a rapid increase from less than 1% earlier in the year.

The disclosure offers a closer look at how AI companies are beginning to use their own models to accelerate the development of more capable systems. It also raises questions about how much of future AI development could eventually be performed by AI itself.

What Anthropic Means by “Leads”

The 26% figure does not mean Claude is independently building a new AI model from beginning to end.

Anthropic's measurement uses an automation scale developed by Epoch AI. On that scale, an AI system that “leads” a task can complete most of the work from a high-level prompt while a human remains responsible for supervision.

That is different from full autonomy. Anthropic says Claude is not operating fully autonomously for any measured subset of its AI research and development work.

The distinction is important because an AI system can perform a substantial amount of technical work while humans continue to define objectives, review results and make important decisions.

Claude's Role Has Expanded Rapidly

Anthropic's measurements show how quickly the company's use of Claude for internal AI development has grown.

The company reported that Claude's share of AI R&D work at the “lead” level reached 26% in August. The comparable figure was below 1% in February, according to Anthropic's measurement framework.

More than 90% of the measured R&D work involved AI at the “collaborate” level or above. In that category, AI can perform substantial portions of a task while working under close human direction.

That means AI assistance is becoming deeply integrated into the development process even where Claude is not independently leading the work.

What AI Can Do Inside an AI Lab

Modern AI research involves many different types of work. Researchers and engineers may need to write and debug software, analyze experiments, investigate model behavior, develop evaluation systems and work with large quantities of technical information.

AI systems can increasingly assist with many of these activities.

The growing role of AI in programming and development is also reflected in broader changes across the technology industry, including the use of AI tools described in AI in Software Engineering.

For a company building frontier AI models, improvements in these areas can potentially shorten the time required to conduct experiments and address technical problems.

The Rise of AI Agents Changes the Equation

One important part of this shift is the growing use of AI agents.

Rather than simply responding to individual questions, an agent can be assigned a larger objective and perform a sequence of tasks using tools and intermediate results.

Anthropic said approximately 30,000 agents were active on its internal research platform at any given time as of August. The company also reported that agent actions were subject to monitoring designed to identify potentially unsafe behavior.

This development connects closely with The Rise of AI Agents, as companies move from using AI primarily as an assistant toward systems capable of carrying out increasingly complex workflows.

Why AI Helping Build AI Matters

The significance of Anthropic's disclosure goes beyond one company's productivity.

If AI models become better at conducting research, writing software, running experiments and analyzing results, they could also help developers create the next generation of AI systems more quickly.

That creates a potentially reinforcing cycle:

1. More capable AI assists researchers.
2. Researchers use that assistance to improve AI systems.
3. Improved systems can perform more sophisticated research tasks.
4. Those capabilities can then be used to accelerate further development.

Anthropic describes the possibility of AI eventually being able to build its own successor without human involvement as recursive self-improvement. The company says measuring AI's contribution to development is one way of tracking how close the industry may be to such a scenario.

The current figures do not show that this point has been reached. Claude remains under human supervision for the R&D work Anthropic measured.

Why Measurement Matters

Anthropic's new R&D Automation Index is designed to quantify how much of its AI development work is being performed by AI.

To build the measurement, Anthropic catalogued different types of R&D tasks and assessed their levels of automation. The company said its researchers used internal work records to identify roughly 15,000 granular tasks from sampled work during July and organized them into a larger hierarchy of AI R&D activities.

The company says it intends to continue publishing measurements so that the pace of AI development can be tracked over time.

There are also limitations. Anthropic acknowledges that it is measuring its own systems and that different AI developers could use different methodologies. Without common definitions and independent verification, comparing automation figures across companies may be difficult.

AI Development and the Safety Question

The faster AI systems become involved in developing future AI, the more important evaluation and oversight become.

An AI system that helps write code or analyze research can produce useful results, but it can also make mistakes. In a high-stakes development environment, errors may affect subsequent experiments, evaluations or engineering decisions.

This is why the industry's growing emphasis on testing is closely connected to the issues explored in AI Safety Testing Becomes Global Focus Following Frontier Model Incidents.

Anthropic reported that all actions by its research and engineering agents pass through an online monitoring system before execution, while an offline monitoring system reviews actions afterward. The company said more than one billion decisions were made by these agents in August and approximately one in 47,000 decisions was blocked.

These figures illustrate the scale of oversight required when thousands of AI agents are performing technical work simultaneously.

How This Could Affect AI Research

AI involvement in research and development could change how technical teams allocate their time.

Instead of spending as much time on repetitive implementation or information-gathering tasks, researchers may increasingly focus on defining problems, evaluating results, designing experiments and deciding which directions deserve further investigation.

That does not necessarily mean fewer humans are needed. It can instead mean that individual researchers are able to oversee larger volumes of technical work.

The effect could be particularly significant in areas where software development, experimentation and analysis can be partially automated.

The Difference Between Assistance and Autonomy

The 26% figure is significant, but its meaning needs to be kept in context.

Claude is helping lead a substantial share of Anthropic's measured AI R&D tasks, yet Anthropic explicitly says the model is not fully autonomous in any measured category. Human supervision remains part of the process.

That distinction separates today's AI-assisted development from a hypothetical system capable of independently setting its own objectives, conducting research, evaluating its discoveries and designing a successor without meaningful human involvement.

The industry has not reached that point based on Anthropic's current measurements.

A New Way to Track the Pace of AI Development

Anthropic's disclosure provides a quantitative view of something that has often been discussed more generally: AI companies are increasingly using AI to build better AI.

The reported 26% share of AI R&D led by Claude shows that this transition is already substantial inside Anthropic, while the company's finding that more than 90% of measured work involves AI collaboration or greater involvement points to an even broader role for AI across its development process.

The more important question now may be how these numbers change over time. Continued measurement could show whether AI involvement in model development grows gradually, accelerates sharply or reaches practical limits.

For now, Anthropic's data provides a snapshot of an industry in which the tools used to develop artificial intelligence are increasingly becoming part of the development process themselves.