AI systems are becoming increasingly powerful and are beginning to automate more of the process of building themselves. Anthropic said this acceleration could make it more challenging for humans to understand or control these systems.
Anthropic’s latest measurements show how far that automation has progressed inside the company. As of August 2026, Claude was at the AL4 “leads” level for 26% of Anthropic’s model R&D work, up from 1% in March, according to the company’s prototype R&D Automation Index.
Anthropic said Claude was not operating fully autonomously for any measured subset of model R&D, while more than 90% of the work was at or above the AL3 “AI collaborates” level.
Under Anthropic’s definitions, “leads” means Claude can complete most of a task end-to-end from a high-level prompt while a human supervises. Full autonomy is classified separately as AL5.
The R&D Automation Index measures how much of Anthropic’s AI model R&D is performed by Claude. The automation scale, developed by Epoch AI, ranges from AL0, with no AI involvement, to AL5, where work is fully autonomous with no human in the loop.
Anthropic also measured how its growing use of AI agents is monitored. Around 30,000 agents were carrying out research and engineering tasks at any given time on its most-used internal platform in August.
Anthropic said 100% of these agents’ actions pass through online monitoring before execution. The company analysed more than one billion decisions from its research and engineering agents in August and found that 0.002% were blocked, or about one in 47,000 decisions.
Its offline monitor ingests all agent activity after actions are taken. Anthropic said it flags roughly 1- 2 transcripts per 1,000 for further review. About 100,000 transcripts are flagged each week, with around 50 (highest-priority) per week escalated to human reviewers.
Anthropic’s third measurement looks at the compute powering model development. Between July 13 and 20, about 6% of compute used for model R&D was allocated to safety, while safety accounted for about 12% of compute used for AI-driven model R&D. Anthropic said compute share is an imperfect proxy for the amount of safety work being done. being done.
Anthropic said the figures are a snapshot from inside the company and could shift if frontier labs coordinate on the pace of AI development, as CEO Dario Amodei has called for.
Also Read: No OpenAI IPO in 2026; What Happens to Anthropic’s IPO?
