Anthropic Calls for Verifiable Global Pause on Frontier AI
In a blog post, Anthropic warned that AI systems are approaching recursive self-improvement. At this stage, models could autonomously design, train, and enhance their own successors with minimal human input.
Internal metrics are striking. Over 80 percent of code merged into Anthropic’s production systems is now written by Claude. Engineer productivity has increased roughly eightfold. In one case, Claude boosted code performance from 3x speedup in May 2025 to 52x by April 2026.
Benchmarks confirm the trend. – AI task horizons are doubling every few months, and complex coding benchmarks are nearing saturation.
Anthropic highlights major benefits: self-improving AI could drive breakthroughs in science, medicine, and technology. Yet risks are serious. Once models build successors, alignment, oversight, and control grow far more difficult as flaws compound across generations.
The company rejects an immediate unilateral slowdown, which would cede ground to competitors. Instead, it urges leading labs to create a coordinated, verifiable pause system with reliable verification, safeguards against bad actors, and clear resumption rules.
Anthropic says it would join if the framework is robust and widely adopted.
The firm plans further research and discussions with policymakers and peers to develop practical pause options. Creating a credible brake for frontier AI may be one of the defining governance issues ahead.
