Anthropic has just released a suite of tools to measure three core aspects of AI development: the extent to which AI builds its next version, the ability to monitor autonomous agents, and the amount of compute allocated to safety versus capabilities. This is the first time a frontier lab has publicly released internal metrics at such a detailed level, and Anthropic is proposing that other labs do the same.

Claude ‘leads’ 26% of R&D work, up from under 1% in 6 months

Anthropic built a metric called the R&D Automation Index, based on Epoch AI’s 6-level rating scale: AL0 (no AI), AL3 (collaborative AI under close human guidance), AL4 (AI-led, completing most tasks from a high-level prompt), and AL5 (fully autonomous, with no human in the loop).

Figures as of August 2026 show: Claude has not reached AL5 for any category. However, Claude has “led” 26% of Anthropic’s R&D work, a sharp increase from under 1% in February 2026. Moreover, more than 90% of tasks are at the “collaborative AI” level or above.

To build this index, Anthropic randomly sampled 20% of employees each week during July 2026, used a Claude research agent to extract tasks from Slack and internal documents, creating approximately 15,000 detailed R&D tasks organized into a 542-node hierarchical tree. Each node was evaluated for its level of automation by an independent Claude judge, which was then cross-referenced against employee evaluations: the model and humans agreed in 59% of cases, and differed by at most one level in 97% of cases.

30,000 agents working simultaneously at Anthropic

The second part of the announcement focuses on agent monitoring. As of August 2026, there are approximately 30,000 agents performing research and engineering work at Anthropic simultaneously. Every action taken by these agents passes through two monitoring layers:

Online monitors inspect actions before execution in real time, usually within seconds. Block rate: 0.002%, meaning roughly 1 out of 47,000 decisions is blocked. This figure is based on an analysis of over 1 billion decisions in August 2026.

Offline monitors review activity post-execution, detecting slow-developing and reversible dangerous behaviors. Each week, approximately 100,000 records are flagged, of which about 50 top-priority records are escalated for human review. Flag rate: 1–2 per 1,000 records.

Anthropic stated that METR, an independent evaluation organization, has red-teamed their offline monitoring platform. The company is also setting up third-party evaluations with access to processes, systems, and internal data equivalent to its internal risk evaluation team.

6% of R&D compute dedicated to safety

The third part addresses compute allocation. During the week of July 13–20, 2026, approximately 6% of AI R&D compute was allocated to safety work, and 12% of AI-driven AI R&D compute was allocated to safety. Anthropic emphasized that this is a conservative estimate: if a token serves both capabilities and safety, it is not counted toward the safety metric.

The value of this metric lies not in absolute numbers but in enabling comparisons across labs and over time. Anthropic proposes that labs publicly disclose their safety compute ratio, with published category definitions and third-party verified classifications.

Context: Dario Amodei calls for ‘pacing the frontier’

This announcement comes amid a broader context: CEO Dario Amodei has publicly called for “pacing the frontier,” meaning slowing the pace of frontier AI development to ensure safety. Anthropic is proposing this as an industry standard: every frontier lab should publicly disclose these three metrics, enabling third-party and government oversight.

The most notable aspect of the publication is not the specific numbers, but the methodology. Anthropic details how the metrics were constructed, evaluated, and cross-referenced with human assessments so that other labs can adopt them. The company also acknowledges limitations: the rating scale relies on a model judge, the task basket is frozen at the sampling time, and agent monitoring has only been operational for a few months rather than years.

For users and outside observers, this signals that the AI industry is beginning to establish its own transparency standards, even if voluntarily. The remaining question is whether other labs will agree to disclose similar data, and whether governments will make these metrics mandatory requirements or leave them as recommendations.