Anthropic on September 17 published a set of prototype measurements designed to show how much of its own research and development is now performed by AI. The headline number: as of August 2026, Claude "leads" roughly 26% of the company's AI R&D work, up from under 1% in February.
[1]The figure matters because it is the first attempt by a frontier lab to quantify, on the record, how close the industry is to AI building AI. The measurement uses Epoch AI's automation scale from AL0 (no AI involvement) to AL5 (fully autonomous, no human in the loop). The 26% corresponds to AL4, where a model carries most of a task from a high-level prompt while a human supervises at key decision points. More than 90% of tasks now sit at or above AL3 ("AI collaborates"). Anthropic says no measured subset of its R&D has reached AL5.
The index was assembled from the bottom up. During each week of July, a Claude research agent reviewed the Slack messages and internal documents of a random 20% sample of staff across every department in the model-building loop, producing roughly 15,000 granular tasks organized into a frozen tree of 542 nodes, weighted by person-hours. Anthropic is candid about the circularity: Claude collected the material and a separate Claude judge assigned the automation levels. In an internal validation, model and human raters agreed exactly 59% of the time; two human raters agreed only 35% of the time.
[1]A second set of measurements covers oversight. In August, roughly 30,000 research-and-engineering agents ran at any one time on Anthropic's most-used internal platform, making more than a billion decisions; the online monitor blocked 0.002% of them, about one in 47,000. Offline monitors flag roughly 100,000 transcripts per week, with about 50 highest-priority items escalated to human review. On compute, a July 13-20 snapshot showed about 6% of compute going to AI R&D was allocated to safety work, rising to 12% when restricted to AI-driven AI R&D; the company calls both estimates deliberately conservative, since work that advances capabilities and safety at once was counted as capability R&D.
[1]The release lands days after CEO Dario Amodei called for "pacing the frontier," and one day after OpenAI disclosed its own framework for reporting model misalignment. Two design choices in Anthropic's internal scaffold make the oversight numbers legible: each agent carries an individual identity that persists across model upgrades, and agents communicate through a shared, open messaging system cross-referenced with transcripts, so monitors can trace behavior across agent-to-agent exchanges. Anthropic says it will embed independent third-party evaluators from multiple organizations and keep publishing the measurements, which were co-authored by Marina Favaro and Phillie Wright with research direction from Jack Clark. METR has previously red-teamed its offline monitoring platform. The limits are real: the task basket is frozen and may miss new kinds of work, the judge is Anthropic's own model, and cross-lab comparison lacks a shared methodology. What the numbers establish is narrower but not trivial: AI-aided R&D has moved from an idea to a measurable, monitored internal production process — with humans still holding the release button.
[1][2]