Anthropic has released a new framework designed to track how fast AI is building itself and how well AI agents are overseen, as CEO Dario Amodei pushes for an industry-wide slowdown.
In a blog post published Thursday, Anthropic outlined three measurements that cover how much of its AI research and development is done by AI, how the agents are monitored, and how computing capacity is allocated.
"As the world considers pacing the frontier, we should do everything possible to minimize the gap between what frontier labs know and what the public knows," the company said in a separate X post. "This means better measuring the development of AI, publishing our findings, and giving society an opportunity to decide how to use this information."
The company said it plans to embed third-party evaluators with access to internal systems and data comparable to what its own risk assessment teams use.
AI-led R&D
Notably, Anthropic reported that Claude "leads" 26% of the company's AI R&D work as of August 2026, compared to 25% in July and 14% in June, indicating that the AI model can complete most of a task "end-to-end from a high-level prompt" while the human supervises.
The company noted that more than 90% of the company's AI R&D work now involves AI collaborating with humans or beyond.
On the second metric, Anthropic said that as of August, it ran roughly 30,000 AI agents doing research and engineering work at any given time on its most-used internal platform. The company said all of the agents' actions passed through an automated monitor system.
Specifically, across more than 1 billion decisions, its online monitors blocked 0.002% of actions - or about 1 in 47,000. Offline monitors also flagged one to two of every 1,000 transcripts for further review.
The company said that from around 100,000 weekly flags, about 50 high-priority cases were escalated. Such blocked actions are supposed to get human review within a week, according to the post.
Compute allocation
Anthropic is also monitoring how compute is split between safety work and other R&D.
The company said it examined its total compute usage during a snapshot week from July 13 to July 20. About 6% of compute used for AI R&D was allocated to safety, while roughly 12% of compute used for AI-driven AI R&D went toward safety.
The Claude developer also noted that it released the framework as a prototype for industry-wide transparency. "We hope to model that transparency by releasing these measurements, and we’ll continue to do so," the company said.
Anthropic's latest framework came after CEO Dario Amodei called for slowing the pace of frontier AI development over the weekend, citing growing safety risks. OpenAI's Sam Altman, Google DeepMind's Demis Hassabis, and xAI founder Elon Musk also expressed similar concerns or backed the idea.
Similarly, OpenAI published a new framework on Wednesday for reporting AI model misalignment and disclosed six case reports of unintended behavior in models the team has documented over the past six months.
"Examples of misalignment may help identify problems other AI developers might encounter as their systems reach similar capabilities, reveal weaknesses in safeguards, or challenge assumptions about model behavior," OpenAI wrote. "Sharing these findings allows others to investigate the same problems, test our explanations, and improve mitigations."