Anthropic published three metrics on Thursday meant to let outsiders track how fast frontier AI development is actually moving, arguing that the gap between what labs know internally and what the public knows is itself a safety risk. The first metric estimates that AI now leads roughly 26% of Anthropic’s own research and development work, with more than 90% of R&D tasks running at or above the level where AI collaborates directly with human researchers.
The second and third metrics look at oversight and resourcing rather than raw capability. As of August, Anthropic said it had no measured area of AI R&D where Claude was operating fully autonomously, and roughly 30,000 AI agents were doing research and engineering work at any given moment on the company’s most-used internal platform. A one-week snapshot of Anthropic’s compute use found about 6% of total AI R&D compute went to safety work, rising to roughly 12% within the AI-led share of that R&D specifically.
The release builds on chief executive Dario Amodei’s three-step plan for slowing frontier development, published the preceding weekend, and is explicitly framed as a template. Anthropic published its methodology alongside the numbers and said it hopes OpenAI, Google DeepMind and other labs adopt comparable measures rather than leaving pace-of-development claims to marketing and leaks.
No competitor has committed to publishing matching figures, and Anthropic’s own metrics are self-reported, unaudited, and describe a company that has every incentive to look responsibly paced. Whether transparency catches on as a norm or stays a one-lab experiment depends entirely on whether anyone else agrees to be measured the same way.
