There is something unusual about Anthropic’s approach.
Just a few days after filing its IPO paperwork with the Securities and Exchange Commission (SEC) for a deal valued at over $1 trillion, the creator of Claude wants to ease off the accelerator.
The Anthropic Institute, its internal think tank, published a paper titled “When AI Builds Itself,” signed by Marina Favaro and Jack Clark, cofounder of the firm, calling for the creation of a global framework to slow down or suspend the development of the most advanced AIs.
The central thesis is simple to state but troubling in its implications. AI models are advancing so quickly that they could soon be able to improve themselves without human intervention. This threshold, referred to by the term “recursive self-improvement,” lies at the heart of Anthropic’s argument.
AI that builds itself: a reality in motion
To back up its case, Anthropic does something unusual in an industry not known for transparency: it publishes its own internal data. And the figures are striking.
In May 2026, more than 80% of the code embedded in Anthropic’s codebase was written by Claude. Before Claude Code’s launch in February 2025, this share was almost zero.
The duration of tasks that the models can autonomously complete has roughly doubled every four months.
In March 2024, Claude Opus 3 could complete tasks that would take about four minutes for a human. A year later, Claude Sonnet 3.7 managed tasks of about one hour and a half. By 2026, Claude Opus 4.6 handles tasks of twelve hours.
If the trend continues, AI systems could by 2027 autonomously accomplish in weeks what now takes a human several weeks to complete.
Today, Anthropic engineers are integrating on average eight times more code per quarter than they did between 2021 and 2025. The startup notes that a single engineer now oversees a workload vastly beyond what a human working alone could produce.
That said, this is not a catastrophe scenario.
Anthropic envisions three possible futures.
One in which progress slows down naturally and a second in which productivity gains accumulate but humans retain control over the research direction. In the third scenario, AI systems would become fully capable of designing their own successors, with human roles confined to supervision and verification.
It is this last scenario that motivates the call for a moratorium.
A 2023 petition as precedent
The request from Anthropic is not without precedent.
In March 2023, more than a thousand signatories, including Steve Wozniak, Elon Musk, and Noam Shazeer (a LaMDA contributor at Google and cofounder of Character.AI), launched a call through the Future of Life Institute to pause for at least six months the training of AI systems more powerful than GPT-4, citing “risks to humanity.”
Elon Musk signed this petition despite his own past as an investor in OpenAI.
The letter urged that this pause be “public and verifiable” and that, if the industry could not voluntarily comply, governments should establish a moratorium. The petition did not lead to concrete follow-up. Since then, the situation has evolved dramatically: the models are far more powerful, and the players far more numerous.
Neither Sam Altman at OpenAI, nor the leaders of Google or Meta, have ever joined such initiatives. Their prevailing culture remains that of accelerating competition, not of a coordinated slowdown.
Without its call, Anthropic cites the precedent of the Intermediate-Range Nuclear Forces (INF) treaty, while acknowledging that this framework took decades to establish, and that time is more pressing for AI.
The main difficulty lies in verification. Training runs are far easier to conceal than missile silos. Their resources are versatile, and the incentive to cheat discreetly is enormous since the one who continues while others pause would inherit leadership.
A credible moratorium would also need to specify what would trigger it, what would terminate it, and who would arbitrate disputes.
Anthropic plans to organize discussions in the coming months with policymakers, researchers, and other AI stakeholders to think through these verification mechanisms. “The window to examine these issues together is open, and outsiders to AI companies must be involved in this deliberation,” it writes.
Accusations of “fear marketing” … and a business hypothesis
The proposal is sure to provoke scepticism.
Anthropic has faced criticism from parts of the industry and White House officials who accuse it of exaggerating risks, or even pursuing a “fear-mongering” strategy, in the midst of the race with OpenAI toward a stock market listing.
A more cynical reading of the move deserves mention.
Anthropic finds itself in a favorable preeminent position. Its annualized revenue is on track to hit $50 billion by the end of June 2026, up from $9 billion at the end of 2025. In the enterprise and developer markets, it seems to have gained a meaningful lead over its direct competitors.
In this context, a global moratorium on advancing top-tier AIs would effectively crystallize the gains and thus lock in a ranking in which Anthropic would emerge prominently.
The argument is not unfamiliar in the history of major technology industries. Established leaders have often been the first to call for regulation, not out of disinterest in business, but precisely because it consolidates their advantage. Halting the AI arms race while you are ahead is a way to turn a conjunctural technological lead into a durable revenue stream for future shareholders.
This hypothesis does not necessarily discount the sincerity of Anthropic’s safety concerns. The two motivations can coexist.
The move sits at a particular moment for Anthropic. Since the launch of Claude Mythos, its most advanced model, it has controlled its diffusion to allow for cybersecurity fixes before releasing it publicly.
Yet it asks for others to slow down even more than it has itself. That is the paradox of the situation. As Anthropic itself acknowledges, a unilateral slowdown by a single lab would merely change the identity of the leader, without creating the collective deliberation process currently lacking.
The call for a moratorium only makes sense if it is universal; which, given current commercial and geopolitical rivalries, remains an uncertain horizon.
