Anthropic puts numbers on how AI is helping build AI

Anthropic has published a snapshot of how AI contributes to its research and development, how it monitors internal AI agents, and how it allocates compute to safety work. The figures offer unusually concrete measures of AI development from inside a frontier lab, though they are Anthropic’s own measurements, not an agreed industry yardstick.

Anthropic Watch analysis

What happened

The company says Claude “leads” 26% of its AI R&D work, meaning it can complete most of a task from a high-level prompt while a person supervises. More than 90% of the work is rated at least “AI collaborates”, where AI handles substantial parts under close human direction. Anthropic says Claude was not fully autonomous in any of the measured areas as of August 2026.

Anthropic also says roughly 30,000 agents were doing research and engineering work at any one time on its most-used internal platform. In a snapshot covering 13 to 20 July, it allocated about 6% of compute used for AI R&D to safety work, and about 12% of compute used for AI-driven AI R&D to safety. The company describes those safety figures as deliberately conservative.

Why it matters

These are attempts to measure something usually discussed in broad, dramatic terms: how much AI is already helping develop future AI, and whether oversight and safety work are keeping pace. The figures give readers a starting point for asking questions that a general claim about “AI building AI” cannot answer on its own.

But comparisons between labs will need a shared method. Anthropic says its current automation index uses its own models to assess its systems, which could introduce errors similar to those being measured. Its compute figures also depend on how work is classified as safety or capability research, a distinction Anthropic acknowledges can be difficult to draw.

Our read

This is a welcome move from slogans towards measurements. The especially useful detail is that Anthropic has reported not only AI’s role in R&D, but also a snapshot of safety-related compute. Neither number settles whether the balance is adequate, and neither is independently comparable across labs yet. Still, a public baseline gives outsiders something better to argue over than a foggy promise to “keep safety in mind”.

What to watch

• Whether Anthropic publishes these measurements regularly, as it says it plans to.
• Whether other frontier labs adopt a common method that allows meaningful comparisons.
• Whether independent evaluators verify the measures and can report incidents.
• Whether reported trends lead to concrete changes in how labs allocate compute or oversee agents.

Discussion spark: Should frontier labs be required to publish comparable figures on AI’s role in building AI and the compute devoted to safety, or are those measures too easy to game?

Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.