Anthropic proposes three metrics to measure the pace of AI development

Anthropic published three metrics to measure the pace of AI development within frontier laboratories: how much R&D is already done by AI itself, how well the actions of autonomous agents are supervised and how computing is allocated. The proposal, released by the Anthropic Institute, is part of a growing concern that more powerful systems are being used to build their own next versions. The goal is to make this rhythm visible from the outside — today it is known almost only by those inside.

Why did measuring the rhythm of AI become a priority?

The central argument is feedback. When a sufficiently capable model helps to write code, suggest experiments or evaluate research results, part of the next generation is built by the current generation. This shortens cycles and increases the speed at which the border advances, but also makes it more difficult for regulators, companies and the public to estimate how fast this is happening. No numbers, any debate about security, regulation or strategy is based on intuition. The publication of metrics is an attempt to replace intuition with follow-up.

What are the three AI metrics proposed?

Anthropic describes three dimensions. The first measure as to the R&D work in AI is done by the AI itself — that is, what fraction of research and engineering tasks has already been delegated to automated systems. The second assesses the quality of supervision of the actions of agents, that is, whether humans and verification systems can monitor and correct what an autonomous agent does in real environments. The third deals with computing allocation: how processing power is distributed among training larger models, running experiments and other purposes. The three connect: an accelerated pace of R&D is only sustainable if supervision follows, and supervision depends on how much compute is reserved for it.

Connection with previous Anthropic publications

The initiative is part of a broader line. The Anthropic Institute: Anthropic's new bet on a reliable and interpretable AI was created precisely as a structure to address governance, transparency and risks of advanced systems. The publication of metrics is a practical step within this agenda: instead of just proposing principles, the company offers concrete indicators that other laboratories could adopt. The full text is available at anthropic.com/institute.

What changes for developers in Brazil?

For Brazilian teams that build products on border models, metrics function as a sign of how providers are thinking about their own advancement — and of what capabilities can reach faster than expected. For those who regulate or hire, they offer a common vocabulary: instead of discussing "advanced AI" generically, it becomes possible to ask how much of the development is automated, how the agent is supervised and how the infrastructure is allocated. None of these metrics, however, are mandatory adoption — they depend on each laboratory publishing its own numbers.

What aren't AI metrics working out yet?

The publication defines what to measure, but does not deliver the values. There are no public targets, safety limits or reporting obligations on the part of other laboratories. Nor is it clear how comparable these measurements would be between companies with distinct definitions of "R&D made by AI" or "supervision". Without standardization and external verification, the risk is that metrics see pieces of communication, not instruments of monitoring.

Frequently Asked Questions about AI metrics

What exactly is Anthropic proposing?

Three metrics to track the pace of AI development: how much R&D is done by AI itself, how well the actions of agents are supervised and how computing is allocated.

Why did supervising agents become a rhythm metric?

Because more autonomous agents can only be used safely if humans and verification systems can track their actions — and that capacity is what limits the actual speed of adoption.

Are metrics only valid for Anthropic?

Nope. They were published as an open proposal, but adoption by other frontier laboratories is not yet guaranteed.