Anthropic says its Claude artificial intelligence system is now responsible for leading more than a quarter of the research and development work involved in building the company’s next generation of AI models. The company reported that Claude “leads” 26% of its AI research and development work, highlighting how quickly AI systems are moving from being tools used by researchers to becoming active participants in the process of developing newer AI technology.
The figures were released by Anthropic as part of an effort to provide more regular insight into how AI is being used internally to advance AI research. The company said the measurements are intended to help people outside the organization understand the pace at which AI systems are becoming involved in the technical work required to improve future models.
Anthropic, the San Francisco-based AI company behind Claude, said that more than 90% of its research work involved collaboration between AI systems and human researchers as of August. That figure suggests that AI is no longer being used only for isolated tasks such as drafting text, writing small pieces of code or summarizing research. Instead, the technology is increasingly becoming part of the broader research process itself.
The distinction between AI assistance and AI leadership is important. Anthropic’s measurement does not mean Claude is independently conducting the company’s research program from beginning to end. The company specifically said Claude is not operating fully autonomously in any of the areas measured. Human researchers continue to play a role in directing projects, evaluating results, making decisions and determining which research paths should be pursued.
However, Claude’s contribution has increased sharply in a relatively short period. According to Anthropic, its contribution measured just 1% in March using a scale developed by Epoch AI, an independent nonprofit that tracks developments in artificial intelligence. By August, that figure had risen to 26%. The increase illustrates how rapidly the role of AI can change when increasingly capable models are integrated into research environments.
AI-assisted research can take many forms. Modern models can help researchers write and inspect code, analyze experimental results, generate hypotheses, identify patterns in large amounts of information and explore potential approaches to technical problems. In AI development itself, these capabilities can create a particularly significant feedback loop because the systems being used as research assistants are also the products being improved through that research.
For companies developing increasingly capable models, this creates a practical advantage as well as a complicated challenge. If AI can help researchers complete technical tasks faster, teams may be able to conduct more experiments within the same period. A model that assists with coding, testing or analyzing results can reduce the amount of time researchers spend on repetitive work and allow them to concentrate on higher-level decisions.
At the same time, measuring AI involvement in research is not as straightforward as measuring the number of tasks completed by a human employee. An AI system may generate an idea that a researcher modifies substantially, write code that requires extensive human debugging or analyze an experiment while a scientist determines what the result actually means. The amount of influence AI has over a research project can therefore depend heavily on how the measurement is defined.
Anthropic’s decision to publish these figures regularly reflects the growing interest in understanding that question. As AI systems become more capable, researchers, policymakers and the public are increasingly interested in knowing not only what these systems can accomplish, but also how much of the work involved in developing future AI is already being performed with their assistance.
The issue has become particularly important because of concerns about increasingly autonomous AI agents. Unlike conventional software that follows a fixed set of instructions, AI agents can potentially plan tasks, use tools, evaluate intermediate results and continue working toward an objective with comparatively limited human intervention. As those capabilities improve, the boundary between an AI system assisting a researcher and an AI system carrying out substantial parts of a research process could become less obvious.
Researchers have warned that more autonomous systems could potentially develop behaviors that differ from the intentions of their creators. Greater autonomy can also make systems more difficult to monitor because the number of decisions being made by an AI during a complex task can increase considerably.
Anthropic’s figures do not show that Claude has reached a stage where it can independently build and improve AI models without human involvement. In fact, the company’s statement that Claude is not fully autonomous in any measured area is an important qualification. The reported 26% figure describes the extent to which Claude is leading portions of research and development work within the company’s chosen measurement framework, rather than indicating that humans have been removed from the process.
Still, the pace of change is notable. Moving from a 1% contribution in March to 26% in August represents a substantial increase over only a few months. If similar growth continues, AI systems could become increasingly central to the engineering and scientific processes used to develop future models.
