Anthropic Says 26% Of Its R&D Work Now Reaches Level 4 AI Autonomy
A frontier lab is now publishing its own internal automation numbers, which lets outside evaluators check the pace claims that labs otherwise make about AI developing AI.
Reporting from 1 source: GIGAZINE.
Anthropic published internal metrics on its own AI use. As of August 2026, 26% of its research and development work had reached Level 4 on its R&D Automation Index, where AI carries out most work under broad human instructions. Adding Level 3, the figure exceeds 90%. About 30,000 AI agents ran simultaneously on its main internal agent platform. No measured task reached Level 5, fully autonomous.
The index classifies AI involvement across six levels, from Level 0, where humans do all the work, to Level 5, where AI runs from discovering a task to executing it. Anthropic based the index on evaluation criteria proposed by the research organization Epoch AI. Level 4 means AI carries out most of the work given broad human instructions.
On monitoring, Anthropic combines checks on agent operations before execution with analysis of activity logs afterward. In August 2026 it monitored more than 1 billion decisions, and staff review about 50 high-priority items each week. For the week of July 13 to 20, 2026, about 6% of research and development compute went to safety-related work, rising to about 12% when limited to AI doing AI research.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.