Anthropic Says Claude Now Leads 26% of Its AI Model R&D

Anthropic has disclosed that its Claude artificial intelligence system now leads 26% of the company’s research and development work on AI models, a sharp increase from 1% in March. The company says Claude is not operating fully autonomously, but can complete large parts of certain research tasks under human supervision.

The disclosure provides one of the clearest recent indications of how frontier AI companies are beginning to use their own models to accelerate the development of newer systems. Rather than AI being used only as a product sold to customers, it is increasingly becoming part of the infrastructure through which AI companies conduct research, write code, evaluate models and develop new capabilities.

Anthropic said that more than 90% of its research work in August involved collaboration between humans and Claude. The company also reported that around 30,000 AI agents were carrying out research and engineering work on its internal platform during the month. Anthropic stressed that these systems remain subject to human oversight and that Claude is not currently working completely autonomously.

AI Is Moving Further Into the Development Process

The significance of the 26% figure is less about a single model taking over a company's research operation and more about the changing economics of AI development.

Building frontier models requires enormous amounts of computing power, engineering time and specialised research. If AI systems can increasingly perform substantial portions of those tasks themselves, the productivity of research teams could change significantly.

Anthropic's figures suggest that this transition is already measurable. In February, Claude was not leading any of the company's model R&D work under the metric it now uses. By August, it was leading 26%. The company describes “leading” as completing most of a task from a high-level prompt while humans retain a supervisory role.

That distinction is important. The current process still depends on people to establish objectives, supervise the work and intervene when necessary. The development therefore represents a shift toward human-AI collaboration, rather than evidence that AI research has become fully autonomous.

The Race Is No Longer Only About Better Models

For years, competition between AI companies was largely measured through model performance, benchmark results, computing capacity and access to users.

That competitive equation is beginning to broaden.

The ability of an AI company to use its own models to accelerate research could become an important operational advantage. A system capable of assisting with software engineering, experimentation, evaluation and scientific research can potentially increase the amount of work that a relatively small team can complete.

This creates a feedback loop. Better models can help researchers build better models, which can then be used to accelerate the next development cycle.

Anthropic itself says that measuring this process is important because models accelerating their own development could make it more difficult for humans to understand or control the systems. The company has therefore called for frontier AI developers to publish comparable measurements of their internal development processes.

The Measurement Problem

There is also a limitation to the headline 26% figure.

It does not mean that Claude independently developed 26% of Anthropic's next model. The measurement covers the share of R&D tasks where Claude can perform most of the work while remaining under human supervision. The methodology was developed with input from Epoch AI, an independent nonprofit that studies AI progress, and Anthropic said humans reviewed the results.

That makes the number useful as an indicator of changing AI capabilities, but it should not be interpreted as a precise measure of how much of the next Claude model was “built by Claude.”

The distinction will become increasingly important as AI companies begin publishing more information about how their systems contribute to their own development. Without common definitions and methodologies, comparisons between companies could become difficult.

Safety Becomes Part of the Development Equation

Anthropic's disclosure also comes at a moment when AI safety has become more closely connected to the pace of model development.

The company said that around 6% of its computing resources used for AI R&D in a sampled week in July went toward safety-related work, rising to 12% for research carried out by AI itself. Anthropic noted that these figures may underestimate the amount of safety work because some activities contribute simultaneously to capability and safety.

The timing is significant. OpenAI announced a day earlier that it would introduce a framework for tracking and publicly reporting unexpected or concerning behaviour from its models, following the disclosure of six incidents involving model behaviour over the previous six months.

The two developments point to the same structural issue from different directions: as AI becomes more deeply involved in the development of AI, measuring its contribution and monitoring its behaviour become increasingly important at the same time.

From AI-Assisted Work to AI-Assisted AI Development

The commercial implications extend beyond the laboratories of Anthropic and its competitors.

If AI can substantially increase the productivity of research and engineering teams, the economics of developing advanced models could change. Companies may be able to conduct more experiments, write and test more code, analyse more data and evaluate more model versions without increasing human teams at the same rate.

That could intensify competition among frontier AI companies while also increasing demand for computing infrastructure. The more productive AI research becomes, the more rapidly companies may be able to translate additional computing capacity into new model development.

It also creates a different kind of competitive advantage. Access to capital and chips remains essential, but the ability to use AI effectively inside the AI development process itself could become an equally important capability.

Anthropic's latest disclosure therefore represents more than a new statistic about Claude. It provides a window into a potential change in how the AI industry operates.

The companies building the next generation of AI are increasingly using AI to help build it.

The strategic question is no longer simply how capable the next model will be. It is increasingly about how much of the process of creating that model can itself be accelerated by AI, and how effectively humans can continue to supervise that acceleration.