Anthropic is calling on the world's top AI labs to consider a coordinated slowdown, warning that AI systems may soon gain the ability to improve themselves without human oversight — a threshold it says could arrive within two years.
Back
Anthropic urges global AI pause as models near self-improvement milestone
Anthropic, the $965 billion AI startup behind the Claude chatbot, published a blog post Thursday urging the world's leading artificial intelligence laboratories to consider a coordinated global pause on frontier model development. The company warned that AI systems are approaching "recursive self-improvement" — the ability to design and build their own successors without human intervention — and said that threshold could arrive within two years, or sooner.
"If it were possible to effectively slow the development of this technology to give ourselves more time to deal with its immense implications, we think that would likely be a good thing," Marina Favaro, head of Anthropic's internal research arm, and Jack Clark, the company's head of policy, wrote in the post titled "When AI Builds Itself."
The warning was accompanied by internal data Anthropic had never before disclosed publicly. As of May 2026, more than 80 percent of code merged into Anthropic's production codebase was written by Claude, the company's AI model. The average engineer now merges eight times as much code per day as in 2024. In a March survey of 130 research staff, the median employee estimated that using the company's latest model, Mythos Preview, made them roughly four times as productive as working without any AI tool.
The call for a slowdown arrives at an awkward moment for Anthropic. The company recently completed a $65 billion Series H funding round that pushed its valuation to approximately $965 billion, making it the most valuable AI startup in the world — surpassing OpenAI. It has also secretly filed paperwork for an initial public offering. Annualized revenue has surged from $90 billion at the end of 2025 to an estimated $500 billion by June 2026, according to the company.
The recursive self-improvement threshold
Anthropic's central concern is a concept known as recursive self-improvement — the point at which an AI system can autonomously design, code, and train its successor models without human direction. The company mapped out a progression: early Claude models (2021-2023) required humans to write all code; by 2023-2025, chatbots assisted with snippets; by 2025-2026, coding agents independently wrote and modified entire files; and today, autonomous agents can run code and delegate tasks to other agents. The final stage — a "closed loop" where Claude builds and improves future versions of itself — has not yet arrived, but Clark wrote that he believes it "may happen within two years, or sooner."
The data supports accelerating capability. The duration of tasks Claude can complete independently has been doubling every four months, down from every seven months previously. Claude Opus 3 in March 2024 could handle tasks requiring about four minutes of human work; Claude Opus 4.6 one year later could handle 12-hour tasks. Mythos Preview, Anthropic's latest model, can sustain work for at least 16 hours continuously, according to the research organization METR, which said the model sits "at the upper limit of what METR can measure without introducing new tasks."
On SWE-bench, a standard real-world software engineering benchmark, model scores rose from single digits to near-saturation in two years. On CORE-Bench, which tests whether models can reproduce published research results, success rates climbed from about 20 percent in 2024 to saturating the benchmark within 15 months.
Critics see a regulatory capture play
Anthropic's safety-first posture has drawn skepticism from rivals and industry observers. OpenAI Chief Executive Sam Altman has accused the company of "fear-based marketing," saying last month: "It is clearly incredible marketing to say, 'We have built a bomb, we are about to drop it on your head. We will sell you a bomb shelter for $100 million.'"
David Sacks, a venture capitalist and informal adviser to U.S. President Donald Trump, has accused Anthropic of running a "regulatory capture agenda" — using warnings about existential risk to push for heavy-handed regulations that would restrict lower-cost open-source models and boost demand for Anthropic's proprietary technology.
Ethan Mollick, a professor at the University of Pennsylvania's Wharton School, posted on X that the blog post contained "a bit of navel-gazing, some marketing, and a lot of very sincere beliefs about what Anthropic thinks is likely in the near future of AI."
Anthropic acknowledged the verification challenge inherent in any global pause. "Training runs are far easier to conceal than missile silos," the authors wrote, comparing the effort to Cold War-era arms control treaties. "A meaningful slowdown or pause would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions."
The company said its Anthropic Institute will work with partners to study what a credible verification system would look like, and that it plans to organize broader debates with policymakers and researchers. If such a system can be built, Anthropic said it would be willing to slow or pause its own development — provided other frontier labs verifiably do the same.
What it means for investors
Anthropic's call for a coordinated pause introduces regulatory uncertainty for the entire AI sector at a moment when capital is flooding in. The company's own data — showing 8x engineer output and 4x productivity gains — simultaneously reinforces the investment thesis that AI delivers measurable economic returns. The paradox is that Anthropic is both the beneficiary of AI's acceleration and the loudest voice urging it to stop.
OpenAI is also expected to pursue an IPO soon, and the two companies together represent roughly $2 trillion in combined market value. Any regulatory framework that emerges from this debate could determine which business models — proprietary versus open-source, safety-first versus speed-first — ultimately prevail. For now, the market has not priced in the risk of a coordinated global pause, in part because the enforcement mechanism remains undefined and the incentives to cheat are enormous.
This article is for informational purposes only and does not constitute investment advice.
[1] Rare! "AI Giant" Calls for Global Pause on AI Development, Warns Models Are Gradually Gaining "Self-Improvement" Capabilities[2] Anthropic Urges Global Pause in AI Development, Flags ‘Self-Improvement’ Risk - WSJ[3] Anthropic calls for global pause in AI development before humans lose control[4] Anthropic calls for global AI slowdown after $965B valuation. Critics claim it’s just to hobble competition.[5] Anthropic calls for global freeze in AI development
