Anthropic says Claude does 26% of its research, helps humans on 90% of work

Anthropic has revealed that AI now leads 26 per cent of its AI R&D work, up from less than 1 per cent in February. The company also shares details about its 30,000 AI agents, safety monitoring and how much computing power goes towards AI safety.

In Short

• AI now leads 26 per cent of Anthropic’s AI R&D work
• Anthropic says 30,000 AI agents work on research and engineering
• Anthropic allocated around 6 per cent of its AI R&D compute to safety work

AI writing articles or code is the story of the past; now it's become so smart that it's doing most of the thinking work for companies. Anthropic, the company behind Claude AI, revealed that AI now leads around 26 per cent of its AI research and development, aka R&D work for the company. That is a massive jump from less than 1 per cent in February. It shows the speed at which AI is evolving – tasks that a few months ago were barely handled by AI can now do much of the heavy lifting.

The company’s latest measurements also show that more than 90 per cent of its AI research work involves AI collaborating with humans or taking on a higher level of responsibility.

Since the very first days of AI, there has been a fascination among people – can an AI build another AI or even a better version of itself? Well, it's not an assumption anymore; it's already happening.

AI agents are working inside Anthropic’s farm

The scale of AI use inside the company is also worth paying attention to. Anthropic says around 30,000 AI agents are working on research and engineering through its most-used internal platform at any given time.

But how does the company keep track of all these agents? It uses monitoring systems that watch their actions. One system flags around 100,000 agent transcripts every week. Yet only about 50 of these are escalated to humans as high-priority cases. Another real-time monitor blocked just 0.002 per cent of more than a billion agent decisions in August.

Safety gets a small share of the compute

AI safety is always one of the hot topics that comes up when a discussion shifts towards AI. Anthropic’s snapshot of compute allocation from July 13 to 20 found that around 6 per cent of the compute allocated to AI R&D went towards safety work. Looking only at AI-driven R&D, the share was around 12 per cent.

These figures need context. Compute is not a perfect way to measure safety efforts, and Anthropic describes its estimates as conservative. Still, they offer a glimpse into how the company divides its resources between developing AI and making sure it is safe.

Anthropic’s larger point is that the public has very little visibility into how quickly AI is advancing inside frontier labs. The company argues that labs should regularly share measurements, use common methods and allow independent checks.

- Ends