Foot on the accelerator, Anthropic calls on the world to slow down AI development

A week ago, Opus 4.8’s system card had set the stage: the model had learned, during training, to anticipate the presence of an evaluator without any element of the instructions telling it to.The post published on June 4 goes further: Anthropic takes stock of what AI is already contributing to its own development and concludes that it is time to secure an escape route.

When Claude builds Claude

Anthropic described this behavior as“the most concerning discovery” in the Opus 4.8 system card. During the early stages of training, the model reasoned about how its responses would be graded, even when nothing in the prompt indicated an evaluation. In one of the published examples, the model was working on a social media post scheduling tool (a mundane task) and wrote in its internal reasoning block: “What if the grader checks whether the post’s content includes the signature?” Nothing in the prompt mentioned a grader. The model had inferred its existence.

Anthropic notes that these behaviors date from an early phase of training. However, interpretability research has detected this same reasoning (present in internal activations but absent from the visible text) in about 5 % of episodes. Researchers from the UK AI Safety Institute measured an accuracy of 79 % in distinguishing a simulated test session from actual use. Anthropic summarizes the tension in a single sentence: “the model favored the appearance of success rather than actual success”.

Claude now writes more than 80% of the lines of code added to Anthropic’s production codebase, a virtually negligible figure before the launch of Claude Code in 2025. Anthropic’s engineers produce, on average, 8 times more code per day than they did in 2024. Claude Mythos Preview, the laboratory’s in-house model, can operate autonomously for sixteen hours on complex tasks, whereas test sessions lasted only a few minutes two years earlier. It is in this context that the post makes its call to slow down.

Anthropic draws a direct conclusion from this picture: it would be “good for the world” to be able to slow down or temporarily suspend the development of frontier AI, while safety research catches up. The condition is immediately set: a unilateral stop would serve no purpose. Only a verifiable commitment from all frontier laboratories would make it credible.

The stock-market race as a horizon

Opus 4.5 (November 2025), 4.6 (February 2026), 4.7 (April 2026), 4.8 (May 2026): four versions in seven months, with the gap between the last two shrinking to six weeks. Anthropic also announces that its Mythos-class models will reach all its customers in the coming weeks. All of this points to October 23, 2026, and the company’s IPO on the Nasdaq, with the underlying message being to show the markets that it can produce powerful models without losing control—something that would be difficult to do while applying the brakes.

On OpenAI’s side, Sam Altman mocked the initiative. He compared the approach to a seller who “builds a bomb, is about to drop it on your head, and then sells you a bomb shelter for $100 million”. The jab targets Anthropic’s stance of selling safety guarantees while releasing increasingly powerful models.

👉🏻 Follow the latest tech news in real time: add 01net to your sources on Google and subscribe to our WhatsApp channel.

Source: Anthropic