TL;DR
Anthropic has disclosed that Claude now leads 26% of its internal R&D, up from under 1% in February, with 30,000 AI agents running continuously inside the company, marking a threshold moment where AI is building AI at scale.
What happened
- Claude leads 26% of Anthropic's AI R&D as of mid-2025, a jump from less than 1% in February, a roughly 26x increase in roughly seven months.
- More than 90% of all AI research work at Anthropic now involves AI either collaborating with humans or taking primary responsibility for tasks.
- 30,000 AI agents are active simultaneously on Anthropic's most-used internal research and engineering platform at any given time.
- Safety monitoring flags 100,000 agent transcripts weekly, but only about 50 are escalated to humans as high-priority; a real-time monitor blocked just 0.002% of more than 1 billion agent decisions in August.
- 6% of total AI R&D compute went to safety work in the July 13 to 20 snapshot; among AI-driven R&D specifically, that share rises to 12%.
Why it matters
- The speed of capability transfer is the headline: going from 1% to 26% AI-led R&D in seven months suggests frontier labs could reach majority-AI-driven research within a year, compressing the timeline for recursive self-improvement.
- 30,000 concurrent agents is an operational scale most enterprises have not approached, signaling that agentic AI is no longer experimental inside top labs.
- Safety compute at 6 to 12% of R&D allocation will draw scrutiny: critics will argue the share is too small given the stakes; Anthropic itself calls the estimates conservative and notes compute is an imperfect proxy.
- Anthropic is making a transparency argument: by publishing these metrics, the company is implicitly pressuring rivals to disclose equivalent data and pushing for common measurement standards and independent audits.
- Competitive signal for enterprises: if a frontier lab is running 30,000 agents internally, the tooling, reliability, and cost curves needed for large-scale enterprise deployment are likely closer than public benchmarks suggest.
What to watch next
- Whether OpenAI, Google DeepMind, or xAI publish comparable internal metrics, which would validate or challenge Anthropic's framing and reveal whether 26% AI-led R&D is an outlier or an industry norm.
- The trajectory of the 26% figure: if it continues at the same pace, AI-led R&D could cross 50% by early 2026, a threshold that would force regulatory and governance conversations about autonomous research systems.
- Policy response to the safety compute disclosure: 6% is a number regulators and safety researchers will cite; watch for it to appear in EU AI Act implementation discussions or US executive-branch AI governance guidance.
Originally published on Present of AI, a daily source-linked AI news timeline. Read the full timeline or browse the open dataset.