Anthropic publishes measures of AI research automation and oversight
Anthropic proposes public measures of research automation, agent oversight and compute allocation, with an initial internal snapshot.
AnthropicAnthropic proposes tracking AI research automation, agent oversight and compute allocation to inform frontier pacing. Its internal measurements are company-reported, with independent verification proposed. Anthropic
Automation and oversight
For August 2026, Anthropic estimates Claude led 26% of measured R&D work under human supervision; no measured category was fully autonomous. Claude helped construct and rate the index. Anthropic warns that model evaluators could share the errors of systems they assess. Anthropic
Anthropic reports roughly 30,000 concurrent research and engineering agents on its most-used internal platform. Its oversight measures cover that platform only. Monitoring coverage and escalation rates do not establish how reliably harmful behavior is detected. Anthropic
Safety compute
During July 13–20, about 6% of AI R&D compute went to safety. This one-week snapshot cannot establish a trend, and compute expenditure is an imperfect proxy for safety effort. Anthropic
Sources & attribution
- Analysis 17 Sept 2026Measurements for understanding the pace of AI development inside frontier labs
Anthropic.