Anthropic is proposing a new way to measure the pace of frontier AI development. In a September 17 announcement, the company argued that outside observers cannot currently see enough of what happens inside leading AI labs, and suggested publishing comparable indicators.
What the proposal is trying to measure
Anthropic’s paper focuses on measurements such as the amount of compute used, the time and cost needed to train systems, and the capabilities those systems demonstrate. The company presents these as a way to make progress easier to compare across time and between organizations.
The proposal is about visibility, not a claim that one metric can summarize an AI system. Training budgets, evaluation results and deployment choices each reveal different parts of development. A number can also be difficult to compare when labs use different tests or keep methods private.
Why transparency matters
Anthropic says better measurement could help researchers, policymakers and the public understand whether progress is accelerating, plateauing or changing direction. It could also make discussions about safety and governance less dependent on occasional product announcements.
Our reading: the proposal is useful only if the industry can agree on definitions and publish enough underlying detail to audit the results. Voluntary metrics would be a start, but independent checks and consistent reporting would determine whether they become a durable public signal.
Anthropic is asking for feedback as it develops the framework. The full proposal is linked below.