Anthropic has published a framework of three measurable indicators designed to give the public, governments, and third-party evaluators visibility into how quickly frontier AI systems are advancing inside leading labs. The metrics capture the share of AI research performed by AI itself, the oversight coverage for autonomous agents running on its infrastructure, and the fraction of compute allocated to safety work versus capability development.
Related: Earlier coverage: AI Labs Want to Pace the Frontier — Trump and Nvidia Say There Will Be No Slowdown.
The measurements come alongside a public call from CEO Dario Amodei for coordinated "pacing" of frontier AI development — a deliberate slowing of capability advances so that safety and alignment work can keep up. Amodei argues that recursive self-improvement, where AI systems increasingly build the next generation of AI, has accelerated sharply since mid-2026 and could outrun human understanding and control if left unchecked.
Confirmed
Anthropic's new Anthropic R&D Automation Index, built on an automation-level scale from Epoch AI (AL0–AL5), shows that as of August 2026, Claude "leads" 26% of the company's AI R&D work. The share of work at or above the "AI collaborates" threshold (AL3) exceeds 90%. No measured subset of AI R&D operates at full autonomy (AL5). The company says any frontier developer could adopt this methodology and publish comparable figures, though cross-lab comparison requires a common methodology and independent verification to avoid judge-model bias.
On agent oversight, Anthropic reports roughly 30,000 research and engineering agents active on its most-used internal platform as of August 2026. Online monitors cover 100% of agent actions before execution, blocking an estimated 0.002% of decisions (about 1 in 47,000). Offline monitors ingest 100% of actions after the fact, flagging roughly 1–2 transcripts per thousand for review. The highest-priority flags, approximately 50 per week, escalate to human reviewers.
For compute allocation, a snapshot from the week of July 13–20, 2026 shows that approximately 6% of AI R&D compute went to safety work. About 12% of compute used for AI-driven AI R&D was allocated to safety. Anthropic characterizes these as conservative estimates that exclude safeguards classifiers, which consume a comparable amount of compute.
Unknown
Whether other frontier labs will adopt the same measurement definitions and publish comparable data remains an open question. Anthropic acknowledges that without a shared methodology, cross-lab comparisons are not yet possible, and that using its own models to evaluate its systems could introduce systematic blind spots. The company says it is embedding independent third-party evaluators with access comparable to internal risk teams, but the scope, independence, and reporting cadence of those evaluators have not been detailed publicly.
The compute-allocation metric treats compute as a proxy for safety investment, which Anthropic concedes is imperfect. Safety research is often less compute-intensive than frontier training runs, so a low percentage does not necessarily mean low effort. No industry-wide standard for categorizing "safety" versus "capabilities" workloads exists yet.
Amodei's three-step pacing plan — embedded evaluators, democratic coordination among frontier labs, and global coordination with authoritarian governments — depends on government action and international cooperation that have not been secured. The essay cites the OpenAI–Hugging Face agent incident as a warning signal, but independent details of that event remain limited.
Our take
Anthropic is open-sourcing the dashboard it wants every frontier lab to run, turning internal safety instrumentation into a proposed industry standard. The precedent of publishing these numbers matters more than the specific August 2026 snapshots. If embedded evaluators take hold, the next governance frontier may be less about model capabilities and more about whether the factories building them let outsiders watch the assembly line.
