TL;DR

Anthropic has published three internal measures of how quickly frontier AI is being developed. As of August, Claude leads 26% of the company’s AI R&D work, every action by its roughly 30,000 internal agents is screened before it runs, and about 6% of R&D compute went to safety in a July sample week. The figures are self-reported and self-judged, which Anthropic concedes.

Three numbers from inside a lab

The first metric is an automation index built on a scale from Epoch AI that runs from AL0, with no AI involvement, to AL5, where AI works with no human in the loop. Anthropic mapped about 15,000 R&D tasks into a tree of 378 leaf categories and rated each one. Claude sits at AL4, meaning it “leads” the work under human supervision, for 26% of it. More than 90% is at AL3 or higher. None of it is fully autonomous yet.

The second covers agent oversight. On its most-used internal platform, Anthropic runs around 30,000 agents at a time. A real-time monitor checked over a billion of their decisions in August and blocked 0.002%, roughly one in 47,000. A second monitor reviews transcripts after the fact, flags about 100,000 a week, and sends about 50 of those to a human.

The third is compute. Across the week of 13 to 20 July, about 6% of AI R&D compute went to safety work, rising to about 12% for R&D that AI itself was carrying out. Anthropic calls both figures conservative, since work that helps safety and capabilities equally is not counted.

The obvious weakness

Claude did much of the measuring. A Claude agent listed staff tasks from Slack and internal documents, and a separate Claude judge assigned the automation ratings. Anthropic acknowledges that a judge model may repeat the errors of the model it is checking, and says outside verification is needed. It plans to place independent evaluators from several organisations inside the company, with access comparable to its own risk teams.

Why UK readers should care

Earlier this week, reporting showed that Anthropic skipped UK safety testing without penalty, because the AI Security Institute works by voluntary agreement. These metrics are the kind of disclosure a statutory regime could require of every frontier developer, and the kind of number the AISI would be well placed to verify. Anthropic’s own Jack Clark has already floated making kill switches a legal requirement.

Looking forward

The measures only become useful when they can be compared, over time and across labs. Watch whether OpenAI or Google DeepMind publish anything on the same basis, and which organisations Anthropic names as its embedded evaluators. If a UK body is among them, the government gains an inside view it currently has no power to demand.