Anthropic has published a proposed set of public measurements for tracking the pace of frontier AI development, arguing that as the world considers slowing the frontier, the public needs visibility into what is happening inside AI labs. The company's post, accompanied by a snapshot of its own internal metrics, is the most detailed self-portrait any frontier lab has released of how much its models now build its models. "Today, the world can't see what's going on inside AI labs," the company wrote, framing the measurements as tools that illuminate three critical aspects of development: the extent to which AI is building the next version of itself, Anthropic's ability to oversee and intervene in actions that AI agents take on its systems, and the resources that power the development of more capable models. ## How Much AI Builds AI The headline number comes from the company's new R&D Automation Index, built by cataloguing every kind of AI research work at the firm and rating how automated each task is on a scale from AL0, no AI involvement, to AL5, AI operating fully autonomously with no human in the loop. As of August 2026, Claude is not operating fully autonomously on any measured subset of work, but the model "leads" 26 percent of Anthropic's AI R&D, completing most of the task end-to-end while a human supervises, and more than 90 percent of work involves AI collaboration. The index was constructed from a bottom-up review of roughly 15,000 granular research tasks, sampled from 20 percent of staff across model-development departments for each week of July 2026, organized by a Claude research agent into a tree of 542 categories. The company notes that models accelerating their own development could make systems harder for humans to understand or control, which is why the distance to recursive self-improvement, a model fully autonomously building its successor, is now a measurable quantity rather than a speculation. ## Oversight At The Scale Of 30,000 Agents The second measurement area covers the agents now embedded in the company's own work. As of August 2026, approximately 30,000 agents were doing research and engineering work at any one time on Anthropic's most-used internal platform, all of them funneled through two layers of monitoring. Online monitors check every agent action in real time before execution, blocking dangerous moves such as an attempt to copy model weights out of company systems. Offline monitors review activity after the fact, flagging roughly 100,000 transcripts per week for automated re-review, with about 50 of the highest-priority flags escalated to human reviewers. Across more than a billion agent decisions analyzed in August, 0.002 percent, about one in 47,000, were blocked. The company also plans to embed independent third-party evaluators with access to internal processes and data comparable to its own risk teams. ## Compute, Safety, And The Pacing Debate The third measurement is compute allocation. In the week examined, about 6 percent of the compute going to AI research was allocated to safety work, and about 12 percent of compute on AI-driven research went to safety, figures the company describes as deliberately conservative. The value, Anthropic argues, is less the absolute numbers than the creation of a like-for-like metric that governments could eventually require of every lab, or attach commitments to, in any coordinated effort to pace the frontier. That debate is the proposal's context. Anthropic CEO Dario Amodei has called for coordination on pacing frontier development, and the company notes its numbers would be expected to shift if such coordination took hold. The measurements are designed to make any slowdown, or the absence of one, verifiable from outside the labs. Whether rival labs adopt comparable reporting is the open question, since transparency imposes costs that secrecy does not. But the release resets the terms of the argument: the company urging that the frontier be paced has now published the ruler it proposes to be measured with, and daring the rest of the industry to do the same. ## How The Numbers Were Built The index's methodology is as notable as its figures. Because no individual could list every research task at a frontier company by hand, Anthropic built the catalogue from the bottom up: for each week of July 2026 it randomly sampled 20 percent of staff from every department in the model-development loop, had a Claude research agent review each sampled person's week through Slack and internal documentation, and compiled roughly 15,000 granular tasks. Claude then organized them into a hierarchical tree of 542 nodes, from training and product at the root down to leaves like eval platform defect diagnosis and serving incident postmortems. The company is candid about the limits of self-measurement. Using its own models to evaluate its own systems risks the judge model sharing the blind spots of the model it judges, and the safety-compute figures rely on category definitions each developer could draw generously in its own favor. The proposed remedies are structural: third-party verification, cross-lab comparison of methodologies, and burden of proof placed on the developer to show that work counted as safety really is safety. ## What Happens Next The company says it plans to embed independent evaluators from multiple organizations inside Anthropic, with access to internal systems comparable to what internal risk teams receive, and to keep publishing the metrics so they can be tracked over time. Any frontier developer, it argues, could report the same measures today using public methodology, from the automation index to agent-monitoring coverage to the safety share of compute. The proposal lands in a week when the industry's biggest names have been publicly entertaining a superintelligence slowdown, and when governments are actively weighing what transparency obligations to impose. By publishing numbers that make its own automation level, oversight coverage and safety allocation checkable, Anthropic has converted an abstract debate about pacing into a set of figures someone can actually hold a lab to. The next move belongs to its rivals, and to the regulators deciding whether those figures should ever be voluntary again.