Anthropic Releases Three Metrics for Tracking AI Development Pace

Anthropic has published three concrete metrics designed to help artificial intelligence companies measure and report on their development velocity, providing practical tools days after CEO Dario Amodei proposed an industry-wide slowdown.

The company shared methodologies for tracking AI-led research and development, oversight of AI agents, and compute resource allocation in a Thursday blog post, according to CNBC. The move addresses criticism that Amodei's weekend slowdown proposal lacked implementation details.

Three measurement frameworks

The first metric evaluates whether AI models operate autonomously in research and development tasks. Anthropic determined that its Claude models are "not operating fully autonomously" for any subset of R&D work it measured.

For agent oversight, Anthropic built a system to monitor and intervene in actions taken by AI agents. The company found approximately 30,000 agents conducting research and engineering work simultaneously across its primary internal platform.

The third metric examined compute allocation over a one-week snapshot from July 13-20. Anthropic reported that roughly 6% of its AI research and development compute went toward safety work. When narrowed to "AI-driven" R&D specifically, approximately 12% of compute resources supported safety initiatives.

Transparency as industry standard

"As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows," Anthropic stated in its post. The company positioned these metrics as complements to capability evaluations, giving external parties a "starting point" to assess development pace.

Anthropic emphasized that the metrics focus on how models are built rather than what they can do, and encouraged other organizations to adopt similar measurement practices.

Why it matters

Amodei's slowdown proposal gained backing from OpenAI CEO Sam Altman, SpaceX CEO Elon Musk, and Google DeepMind Chair Demis Hassabis. By releasing specific methodologies rather than principles alone, Anthropic is testing whether transparency mechanisms can enable coordinated pacing without regulatory mandates. The compute allocation data also reveals that safety work consumes a small fraction of total AI development resources at one of the industry's most safety-focused labs—a baseline that could inform policy discussions about appropriate safety investment levels.

The metrics arrive as researchers issue increasingly urgent warnings about AI's potential for harm, making measurement frameworks a potential prerequisite for any industry coordination on development speed.

Details were first reported by CNBC.