Anthropic says Claude now leads 26% of its AI R&D work: How fast is AI development accelerating?

Artificial intelligence is increasingly being used not just to write code or automate routine tasks, but to help develop the next generation of AI systems themselves.

Anthropic’s latest measurements show how quickly that shift is taking place. The company says Claude now “leads” 26% of its measured AI research and development work, meaning it can complete most of a task from a high-level prompt while humans supervise.

The figure was below 1% in February 2026. More than 90% of Anthropic’s AI R&D work now involves AI at least in a collaborative role. However, Anthropic says Claude has not reached full autonomy in any measured area of AI R&D.

What is Anthropic measuring? 

Anthropic has created an R&D Automation Index to track how much of its AI-development work is being performed by AI.

The scale runs from AL0, where AI is not involved, to AL5, where AI performs the work fully autonomously. At AL4, or “leads”, AI can complete most of a task end-to-end from a high-level prompt, with human oversight.

Claude reached the AL4 level for 26% of Anthropic’s measured AI R&D work by August. Anthropic says the measurement is designed to track the growing role of AI in developing AI, rather than simply measuring conventional workplace automation.

The distinction is important: Anthropic is not saying Claude can independently build its successor. Humans remain involved in the process.

Why recursive AI development matters 

Anthropic is particularly interested in tracking progress towards recursive self-improvement — a point at which AI systems could independently contribute to creating increasingly capable successors.

The company says that point has not been reached.

But as AI takes over more of the research, coding, testing and engineering involved in developing frontier models, the development cycle could potentially accelerate. Measuring how much work AI is performing is therefore becoming an important indicator of the pace of AI progress.

30,000 AI agents working inside Anthropic 

Anthropic is also tracking the use of AI agents across its research and engineering operations.

As of August 2026, about 30,000 agents were carrying out work simultaneously on its most-used internal platform. The company says all their actions pass through an online monitoring system before execution, while an offline system also reviews their activity.

Anthropic analysed more than 1 billion decisions made by research and engineering agents in August. About 0.002% — roughly one in 47,000 decisions — were blocked by the online monitor.

The offline system flagged around one to two transcripts per 1,000 for further review, with about 50 high-priority cases a week escalated to human reviewers.

These figures are intended to show how monitoring needs to evolve as AI agents take on longer and more consequential tasks.

How much compute is going into AI safety? 

Anthropic also measured the computing resources devoted to safety.

During one week in July 2026, around 6% of the compute used for AI R&D was allocated to safety work. For compute specifically associated with AI-driven AI R&D, the share was about 12%.

Anthropic describes these as conservative estimates and notes that compute is not a perfect measure of safety investment because some safety research relies more heavily on human expertise than computing power.

Why Anthropic wants the industry to publish similar data 

Anthropic’s broader argument is that frontier AI companies need better transparency around the pace of development.

The company says other AI labs could publish comparable measurements regularly, making it possible to track changes over time. But it acknowledges challenges, including the lack of standardised methodologies and the risk of using AI systems to evaluate AI-generated work.