Anthropic Urges Global AI Pause as AI Self-Improvement Nears

Anthropic, the AI safety company behind the Claude chatbot, published a landmark blog post on Thursday warning that artificial intelligence systems are approaching “recursive self-improvement” — the ability to autonomously design, build, and train their own successors without meaningful human involvement. The company is calling for a globally coordinated mechanism to pause or slow frontier AI development, arguing that society needs time to build safeguards before it loses control.

In a post titled “When AI builds itself”, authors Marina Favaro and Jack Clark of the Anthropic Institute laid out extensive internal data showing that AI is already accelerating the development of AI systems at a pace that outstrips institutional preparedness. “It would be good for the world to have the option to slow or temporarily pause frontier AI development to enable societal structures and alignment research to keep up with the advance of the technology,” they wrote.

The Evidence: AI Building AI

Anthropic’s internal data paints a striking picture of how quickly the landscape has shifted. As of May 2026, more than 80% of code merged into Anthropic’s codebase was authored by Claude, and engineers now ship eight times as much code per quarter as they did between 2021 and 2025. The company’s AP News reported that this surge reflects a fundamental shift in how AI development happens.

The length of tasks AI systems can reliably complete has been doubling roughly every four months. Claude Opus 3, released in March 2024, could handle tasks lasting about four minutes. By May 2026, Claude Opus 4.6 was managing 12-hour tasks. If the trend holds, multi-day tasks could come into range this year, and multi-week tasks by 2027, according to data from METR cited in the blog post.

Benchmark saturation tells a similar story. Models saturated SWE-bench, a standard test of real-world software engineering, in just two years — going from low single digits to 100%. On CORE-Bench, which tests research reproducibility, models went from roughly 20% success in 2024 to saturating the benchmark 15 months later.

In a March 2026 poll of 130 Anthropic employees, the median respondent estimated roughly four times greater output using the company’s Mythos Preview model compared to working without AI assistance. The company also reported that Claude-written code, which was somewhat worse than human-written code in late 2025, is now roughly at parity and expected to be strictly better within the year.

Perhaps most striking, in April 2026, Anthropic published the first demonstration of Claude running an open-ended research project end-to-end in AI safety. The system recovered 97% of the performance gap that two human researchers recovered only 23% of over a week, using roughly $18,000 in compute.

The Coordination Challenge

Anthropic’s proposed solution — a globally coordinated, verifiable pause mechanism — faces enormous practical hurdles. The company acknowledged that training runs are far easier to conceal than missile silos, their inputs are general-purpose, and the incentive to defect quietly is enormous. “Whoever continues while others pause could inherit the lead,” the blog post notes.

A meaningful slowdown would require multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions — and each must be able to verify that the others have actually stopped. The company compares the challenge to Cold War-era arms control treaties like the Intermediate-Range Nuclear Forces Treaty, which took decades to build both the infrastructure and the trust. “We don’t have that long,” the authors wrote.

Questions of Timing and Motive

The call for a pause arrives at a pivotal moment for Anthropic. The company recently raised $65 billion in a Series H round at a $965 billion valuation, surpassing OpenAI’s $852 billion valuation, and disclosed a $47 billion annualized revenue run rate. It has also filed confidential paperwork for an IPO, racing against OpenAI to go public in what could be one of the largest tech IPO cycles in history.

The timing has raised eyebrows. As Fortune reported, critics have previously accused Anthropic of using safety rhetoric as a form of competitive positioning. Trump adviser David Sacks has accused the company of running a “regulatory capture agenda” designed to slow rivals under the guise of responsible AI development — a characterization Anthropic rejects.

Adding to the credibility questions, TIME reported in February that Anthropic had overhauled its Responsible Scaling Policy, scrapping its central commitment to never train an AI system unless it could guarantee in advance that its safety measures were adequate. Chief science officer Jared Kaplan said the company felt it “wouldn’t actually help anyone” to stop training models unilaterally while competitors pressed ahead.

What Comes Next

Despite the skepticism, Anthropic’s internal data is unusually transparent for a private AI lab and provides compelling evidence that the pace of AI development is accelerating. The company’s SiliconAngle reported that Anthropic Institute plans to convene policymakers, researchers, and civil society in the coming months to work through the questions that recursive self-improvement raises.

“The window to investigate the questions together is here,” Favaro and Clark wrote. “And people outside AI companies should be involved in this deliberation.”

In the short term, the debate over AI safety and regulation will intensify, particularly as Anthropic and OpenAI approach their IPOs. In the medium term, if AI capabilities continue accelerating at current rates, the window for meaningful regulation may close within one to two years. And in the long term, recursive self-improvement could fundamentally transform the economy, potentially making human labor non-competitive in knowledge work.

Whether Anthropic’s call is genuine safety advocacy or strategic positioning ahead of its IPO may ultimately matter less than the question it forces the world to confront: if AI systems can build better versions of themselves faster than humans can manage, who — or what — remains in control?