Anthropic is urging frontier AI developers to implement a verifiable, coordinated pause to address risks associated with recursive self-improvement. The organization seeks this temporary halt to allow global institutions, policymakers, and alignment research to keep pace with rapid technological acceleration and prevent a loss of human control.
Anthropic called on AI laboratories to consider a coordinated and verifiable pause in advanced model developments. The company warns that rapid technological acceleration could soon allow systems to achieve autonomous, recursive self-improvement before society establishes the right safety guardrails.
The proposal arises from concerns that human oversight may become a bottleneck as systems begin designing their own successors. “Right now, it [is] like the AI industry has a gas pedal, but it [does not] have a brake pedal,” says Jack Clark, Co-Founder, Anthropic.
What’s Happening?
The acceleration of AI capabilities has intensified global concerns regarding operational control. Research compiled by the Anthropic institute, the in-house research arm of the company, indicates that the length of tasks models can independently execute doubles about every four months.
In March 2024, the Claude Opus 3 model completed software tasks requiring four minutes of human labor. By March 2025, the Claude Sonnet 3.7 model managed tasks requiring 90 minutes, and Claude Opus 4.6 executed 12-hour tasks. If this trajectory holds, according to the Anthropic Institute, tasks requiring days of skilled labor will come within range during 2026, and tasks requiring weeks could be automated by 2027.
Standardized public benchmarks reflect this acceleration. The SWE-bench, a standard test evaluating software engineering by requiring models to fix bugs in open-source codebases, experienced complete saturation within two years. Similarly, CORE-Bench, which evaluates whether a model can replicate published scientific research, shifted from a 20% success rate in 2024 to full saturation 15 months later. In separate assessments, the Mythos model operated continuously for at least 16 hours, placing its performance at the upper limit of external measurement capabilities.
Internal data from Anthropic confirms that these capability gains accelerate the development cycle. Automation is shifting from code suggestions to autonomous execution over extended horizons.
As of May 2026, Claude authors more than 80% of the code merged into the codebase, compared to single-digit percentages before February 2025. Consequently, during 2Q26, a typical engineer merged eight times as much code per day as recorded in 2024, shifting the human role from active programming to high-level system review.
Challenges to Pause Global AI Development
The call for a global nonproliferation framework emerges as Anthropic navigates a critical financial and corporate expansion. The corporation was recently valued at US$965 billion following a massive funding round and has filed confidentially for an initial public offering in the United States. This commercial positioning places the company in direct competition for public market funding with OpenAI Group PBC, which is also advancing toward an initial public offering.
The rapid commercial growth of frontier developers contrasts sharply with the slow pace of domestic and international regulation. In the United States, where the majority of leading laboratories operate, a recent executive order from the Donald Trump administration placed the responsibility on the corporations themselves, requesting voluntary model submissions for cybersecurity testing prior to public release.
This regulatory vacuum has prompted safety-focused actions from Anthropic, which previously barred the US military from utilizing its models for domestic surveillance and autonomous weaponry. This decision resulted in temporary national security blacklisting slated to take effect later in 2026, though recent reports suggest these regulatory tensions are easing.
Market Reaction After Anthropic Call
The proposal to halt development has drawn varied reactions from market analysts and political advisors who question the feasibility and underlying motivations of a global freeze.
Establishing an international verification regime requires absolute cooperation among rival nations, including the United States and China. Unlike physical military infrastructure, decentralized computing resources and private data centers are exceptionally difficult to monitor.
“This would be practically impossible, because the economic and national security stakes are simply too high for any superpower to willingly hit the brakes now,” says Rob Enderle, Analyst, Enderle Group, suggesting that the safety warnings may function as strategic marketing to validate massive valuations and attract continuous investor capital.
Further skepticism focuses on the potential for regulatory capture. David Sacks, venture capitalist and an informal adviser to Trump, publicly criticized the proposal, suggesting that established corporations use existential narratives to invite heavy-handed state regulations that disproportionately eliminate lower-cost, open-source competitors.
Holger Mueller, Analyst, Constellation Research, notes that a development freeze would primarily benefit front-runners by locking in the status quo, thereby allowing Anthropic to maintain and expand its market share within the lucrative business-to-business sector.
The Anthropic institute acknowledges these structural barriers, noting that a unilateral pause by a single laboratory would fail to establish the necessary deliberative frameworks and would merely shift market leadership to less cautious actors. To address these systemic risks, Anthropic intends to spend the coming months organizing international conversations. These forums will bring together researchers, policymakers, civil society organizations, and competing artificial intelligence corporations to explicitly outline the specific triggers, oversight mechanisms, and verification infrastructures required to make a coordinated global slowdown possible.