Anthropic Discloses Claude Now Leads 26 Percent of Model R&D as AI Self-Improvement Accelerates, CEO Amodei Calls for Industry Slowdown in September 12 Essay

Anthropic announced on September 18, 2026 that its Claude AI model now leads 26 percent of the company's research and development work, completing most tasks end-to-end from a high-level prompt while under human supervision. About 90 percent of Anthropic's R&D is done in collaboration with Claude, meaning the model handles large chunks of work under close human direction. The portion of work Claude leads was zero in February 2026 and reached 25 percent by August, a six-month sprint toward recursive self-improvement. The company disclosed it operates approximately 30,000 AI agents for research and engineering as of August and is establishing external third-party evaluators embedded within Anthropic to monitor safety. The disclosure came days after CEO Dario Amodei published a 3,800-word essay on September 12 calling for a global slowdown in frontier AI development, arguing that capabilities are advancing faster than risk-prevention measures can keep pace. Amodei's call followed an Anthropic researcher's resignation the prior week with warnings about threats AI poses to humanity. OpenAI CEO Sam Altman and Elon Musk have supported the slowdown idea, while other tech leaders and President Trump have pushed back. This matters because Anthropic is openly documenting the velocity of AI recursive self-improvement in real numbers, not vague capability claims. The jump from 0 to 26 percent human-supervised autonomy in six months, combined with 30,000 active agents, shows this is no longer a research horizon problem. Amodei's public call for an industry slowdown, backed by Altman and Musk, signals that even frontier labs building these systems are worried about the control gap. For organizations deploying AI, the practical takeaway is that agentic systems are scaling faster than governance frameworks, and the labs themselves are admitting they lack confidence in their ability to maintain alignment at current development speed.