Artificial intelligence is increasingly deeply involved in the development of its next-generation technologies. According to the latest data from the U.S. artificial intelligence company Anthropic, its AI model Claude’s participation in internal AI development is rapidly increasing, with human-machine collaboration becoming an important mode of research and development work.
Anthropic announced on Thursday (September 17) that as of August 2026, Claude has reached a “dominant” level of work, accounting for 26% of the company’s AI research and development work. The relevant data uses the degree of automation standards formulated by the AI research institution Epoch AI, where “dominant” means that AI can complete the main process of a task based on higher-level instructions, with humans responsible for supervision.
This proportion has significantly increased from less than 1% in February 2026. However, Anthropic emphasizes that Claude has not yet achieved fully autonomous operation in all measured AI development work. At the same time, over 90% of AI development work has reached a level of “AI collaboration” or higher, meaning AI can complete a large amount of work under close human guidance.
Anthropic stated that if the current trend continues, the proportion of AI development work reaching the “dominant” level by the end of 2026 may rise to approximately 80%.
In addition to measuring the level of AI involvement in research and development, Anthropic also revealed for the first time the operation of internal AI agents. The company stated that in August 2026, its most widely used internal platform had an average of around 30,000 AI agents engaged in research and engineering work simultaneously.
Anthropic explained that all operations of these AI agents are subject to online monitoring and system review; after analyzing over 1 billion decisions made in August, approximately 0.002% of decisions were intercepted by the system, meaning roughly 1 in 47,000 decisions were prevented. The intercepted operations undergo manual review within a week.
At the same time, Anthropic released the proportion of computing power used for AI security research. Data from a week in July 2026 showed that approximately 6% of the company’s computing power used for AI development was dedicated to security research; in AI-involved research and development, this proportion was around 12%. Anthropic pointed out that computational work contributing to enhancing AI capabilities and security is included in AI development, thus the above proportions are conservative estimates.
Anthropic stated that one of the purposes of disclosing these indicators is to provide external parties with a clearer understanding of the speed of technical development within advanced AI labs and the level of AI involvement in research and development.
The company expressed that currently there is no unified measurement standard among different AI companies, and self-assessment by AI models may lead to biases. Therefore, in the future, third parties could independently verify the relevant data to increase comparability among different companies.
This disclosure comes at a time when the industry is increasingly concerned about the rapid improvement of AI’s autonomous capabilities. Researchers have previously warned that as the autonomy of AI agents increases, their behavior may deviate from the goals set by developers, posing challenges for monitoring and control.
Anthropic’s CEO Dario Amodei recently published an article on his personal website, urging AI enterprises to slow down the pace of improving model capabilities to allow for more time to implement safety measures.
On September 16, Anthropic’s competitor OpenAI also announced the establishment of a new “misalignment report” mechanism to track, investigate, and publicly disclose unexpected or concerning behaviors exhibited by models, and released six related cases for the first time.