Anthropic says Claude now does most of the work on roughly a quarter of the tasks involved in developing future versions of the company’s AI. Its newly published R&D Automation Index puts 26% of its model research and development at the “AI leads” level as of August 2026.
In Anthropic’s framework, “leads” means Claude can take a high-level prompt and complete most of a task end-to-end while a person supervises. The company says Claude is not fully autonomous for any measured category of R&D. More than 90% of the work in its index is at least at the lower “AI collaborates” level, where the model handles large portions under close human direction.
The disclosure offers an unusually specific view into how a frontier AI company uses its own models, but it is still a company-built measurement. Anthropic used Claude agents to map and assess thousands of internal tasks, then compared some ratings with human reviewers. It acknowledges that borderline judgments remain debatable and that comparisons across labs will require a common method and outside verification.
Thirty thousand agents, monitored by more AI
Anthropic says approximately 30,000 agents were doing research and engineering work at any one time on its most-used internal platform in August. It reports that every action passes through online monitoring before execution and offline monitoring afterward.
The offline system flags roughly 100,000 transcripts each week for additional automated screening. About 50 high-priority cases reach human reviewers, according to the company. The scale helps explain why Anthropic is calling for standardized public reporting on AI-led R&D, agent oversight and the amount of computing power devoted to safety work.
A wider push into life sciences
The same week, Anthropic launched its Life Sciences Verification Program, a beta that gives vetted teams and institutions broader access to Mythos, Opus and Sonnet models for biology work. A standard tier covers most research and development workflows. A project-specific high-risk tier requires additional vetting and renewal every six months, while other safeguards remain active.
Anthropic says the program monitors activity against approved uses and retains flagged data for 30 days. It says that data is kept separate from model training and from the company’s life-sciences research teams.
The initiative reaches beyond software access. Reuters reports that Anthropic has established a wet lab in the San Francisco Bay Area, which life-sciences chief Eric Kauderer-Abrams confirmed. The company is exploring how Claude could direct robotic lab equipment with limited human intervention, although it says human oversight remains essential. A spokesperson also told Reuters that the lab is not specifically for drug discovery, and Anthropic says it is focused on preclinical research rather than running clinical trials.
Tech