OpenAI and Anthropic want to slow down AI, but only if the entire industry slows down with them

OpenAI and Anthropic have been sharing a fascinating and frankly unsettling story for the past few days: the tale of an artificial intelligence that may soon be able to improve itself independently. The most disturbing aspect isn’t the scenario itself, but the fact that those discussing it continue to accelerate development at the same time.

The Shift in Research Labs: From Promises of Acceleration to Public Warnings

Since early June 2026, a term has repeatedly surfaced in the writings of OpenAI and Anthropic: recursive self-improvement. The concept sounds straight out of a science fiction novel, yet it is clearly articulated in their texts. Anthropic considers it plausible, while OpenAI aims for systems capable of significant breakthroughs by 2028 and beyond.

This shift in vocabulary changes everything. For a long time, consumer AI was marketed as a more practical and faster assistant than search engines. Now, industry leaders are describing a future in which models might assist in designing their successors. This shift transcends technical debate and transforms the race for AI into a question of civilization.

A Global Race Where Slowing Down Is Unthinkable As Rivals Accelerate

Here lies the paradox. Anthropic has called for a global coordination system that would allow for a temporary halt in the development of advanced AI if the risks escalate too quickly. OpenAI advocates for multilateral coordination as well, suggesting that slowing down development should be an option when the situation demands it.

On paper, this notion seems reasonable; in reality, it resembles a game of poker. Everyone would reveal their cards only if no one, be it Google, a Chinese competitor, or any other lab, capitalizes on the opportunity to get ahead. A slowdown becomes acceptable only if it is universal, verifiable, and simultaneous. This is precisely what is lacking today.

The most revealing detail lies elsewhere. Anthropic has revised how it presents its safety commitments. The company now differentiates between what it promises to implement alone and what it recommends for the rest of the sector. The message has become more nuanced yet also more brutal: yes to slowing down, but without risking a fall behind the competition.

Increasingly Powerful Models Released Despite Calls for Caution

This tension bursts into the open with the launch, and subsequent suspension, of Claude Fable 5 and Claude Mythos 5. Anthropic introduced Fable 5 as a version sufficiently constrained for general use, while simultaneously acknowledging serious cybersecurity risks associated with Mythos. The situation escalated even further with the swift suspension of access to these models.

The contrast is striking. On one hand, labs describe systems that can help identify software flaws or expedite sensitive tasks. On the other, they continue to release new models at a rapid pace, nearly monthly. The displayed caution seems less like a halt and more like driving under foggy conditions.

IPO, Open Source, and the Battle for Influence Over Real Risks

This double talk occurs at a particularly critical moment. In early June, Anthropic and OpenAI confirmed plans related to a potential IPO. In this context, discussions on safety may stem from genuine conviction or serve to demonstrate to investors and regulators that there is a responsibility doctrine concerning a technology that has become politically sensitive.

Critics emerge from opposing camps. Some pro-innovation officials accuse Anthropic of playing with fear, arguing that the company pushes for regulations that would predominantly disadvantage open source models. Conversely, Gary Marcus and Meredith Whittaker contend that futuristic scenarios obscure already present dangers.

Perhaps that’s where the true vertigo lies. The debate is no longer solely about the possibility of uncontrollable AI. It focuses on who has the right to define risk, when, and with what interests behind the scenes. When the fastest companies call for a slowdown, it is never just a simple alert. It is also a struggle to determine who will take the wheel.

Scroll to Top