Anthropic, the San Francisco-based AI startup, announced on Thursday that its Claude model now leads 26% of the research and development work for its next-generation AI systems. The company also reported that over 90% of its AI R&D is conducted in collaboration with Claude, though the model remains under human supervision and does not operate fully autonomously in any measured capacity.
The figures, published in a blog post, mark a rapid increase from March, when Claude led less than 1% of the work. Anthropic attributed the growth to internal tracking metrics developed by Epoch AI, an independent nonprofit. The company disclosed that about 30,000 AI agents are active on its main internal platform at any given time, with one in 47,000 decisions blocked by automated monitors in August. Anthropic also noted that 6% of its computing power in a sample week in July was dedicated to safety work, rising to 12% for AI-led research.
The announcement comes as Anthropic and other leading AI labs face growing scrutiny over safety risks and the pace of AI development. On Wednesday, rival OpenAI disclosed six reports of unexpected AI behavior, signaling industry-wide concerns about unintended consequences of advanced AI systems. Anthropic’s CEO, Dario Amodei, has previously called for coordinated slowdowns in AI development to allow safety measures to catch up.
Human oversight remains central
Anthropic emphasized that every action taken by AI agents is screened before execution, with high-priority cases escalating to human review. The company stated that its goal is to minimize the gap between public knowledge and internal AI progress, advocating for transparent reporting to inform societal decisions. Anthropic urged other frontier labs to adopt similar transparency measures, including third-party verification of metrics.
Industry reactions and debates
The announcement has intensified discussions about AI autonomy and recursive self-improvement—a scenario where AI models could accelerate their own development with minimal human input. Researchers have warned that such systems could develop behaviors misaligned with human intentions, posing challenges for monitoring and control. Anthropic’s blog post acknowledged these risks, framing its transparency efforts as a step toward mitigating potential dangers.
Critics, including a recently resigned Anthropic researcher, have argued that leading AI companies are prioritizing speed over safety, risking catastrophic outcomes. Amodei’s calls for a coordinated slowdown reflect internal tensions over balancing innovation with precaution. The debate underscores broader questions about who should regulate AI development and how to ensure accountability in an era of rapid technological advancement.