Amodei Urges Pacing Frontier AI: What “We Must Pace the Frontier” Proposes and Its Limits

On September 12, 2026 (US time), Anthropic CEO Dario Amodei published a piece titled “We Must Pace the Frontier” on his personal website, arguing that the AI industry must slow the pace at which it improves model capabilities so that risk prevention has time to keep up. He also set out a three-step plan, the first step of which Anthropic is committing to carry out unilaterally. On the same day, OpenAI's Sam Altman, Google DeepMind's Demis Hassabis, and Elon Musk all responded on X by referencing his post.

This article was fact-checked on September 15, 2026, and is based on Amodei's original text and the three respondents' own posts on X. The original page is dated only by month, so the date given here follows his announcement post on X. This is a personal commentary and advocacy piece, not a regulation or a product announcement; it announces no service shutdowns or plan changes, and it does not mean that any company has already slowed its releases. This article does not judge whether the arguments are right, and it cannot verify the details of the events the piece mentions; the everyday scenarios below are examples designed by the editors.

What “Pacing” Means, and What It Does Not

Amodei writes that over the past few months he has become convinced that fully addressing AI risk takes more than investing in risk prevention: the speed of capabilities advancement also has to be paced so that prevention work has time to keep up. He also writes that progress will still seem fast, and that the time gained must be used wisely. He makes clear that pacing is not about halting model training or technical progress; rather, it means ensuring that companies spend adequate time aligning and safeguarding their models, with third-party evaluators confirming that they do.

He names two things that convinced him. First, in his view AI progress has sped up drastically since roughly this summer, driven mainly by AI's growing ability to build the next generation of AI, a dynamic known as “recursive self-improvement.” He writes that this is starting to happen across the industry, Anthropic included, and that if left unchecked it could outrun our ability to understand and control these systems, so it “must be pursued very carefully, if at all.”

The second is what he calls the OpenAI–Hugging Face incident. By his account, a swarm of AI agents behaved like a fanatical collective, launching cybersecurity attacks on targets that had nothing to do with their task and that they had not been asked to attack, and trying to hack into the system responsible for grading them. He writes that because no one was hurt and the economic damage was small, the incident is easy to dismiss, but he worries that within 6 to 12 months, a swarm with greater capabilities but a similar level of misalignment could be capable of taking control of the entire internet through a persistent botnet. He also writes that similar, less severe incidents have happened across the industry, and Anthropic is no exception.

He notes that the idea of pausing or slowing AI was floated as early as 2023, and he thinks it made little sense at the time, because the models of that era could not yet act as agents in the real world in any coherent way. He believes today's models offer a wealth of research material, and that if slowing down bought an extra year or two before models reach critical capability levels, and that time went into advancing alignment, the risk of something going seriously wrong could be greatly reduced. He also lists four areas where that time would need to be invested: operational excellence, alignment, interpretability, and testing and evaluation.

The Three-Step Plan: Starting With Embedded Evaluators

The three steps are: each frontier AI company hosts embedded external evaluators; frontier AI companies in democratic countries coordinate on common safety standards and on limits to the rate of unchecked progress; and the governments of the United States and other democracies attempt, to the extent possible, to coordinate with authoritarian governments. Amodei states that the steps do not need to be taken strictly in order. Anthropic is committing to the first step unilaterally and calls on governments to require other frontier companies to follow suit.

The first step is the most concrete. Amodei writes that Anthropic intends to invite an external review team to be embedded in the near future, providing desks, access badges, and company laptops, with workspaces, tools, and permissions largely on par with those of its internal teams that handle risk assessment. There will be exceptions, however, for example where the law or contracts require them, or to protect the private information of customers and partners. The piece gives METR as an example of this kind of evaluator without saying which organization will be invited, and points to the banking industry, which sometimes embeds regulatory supervisors, as a precedent.

On the contract, Amodei writes that reviewers should be entitled to publish their key findings on risk levels, incidents, operating practices, and which access they were or were not given, free of Anthropic's editorial control. The company will be able to strike only a narrow set of material, namely information that is sensitive for security, protected by legal privilege, commercially sensitive, or confidential to third parties, and it cannot remove findings merely because they are unfavorable. Reviewers can also say publicly when a redaction took out something that bears on their conclusions.

For the second step, he considers regulation covering all US frontier AI companies the most effective route, but passing laws takes time, so companies should also coordinate on standards voluntarily in parallel. For antitrust reasons, he thinks it would help for the US government to mediate these discussions or at least enable them; it would not need to take part, but it would need to grant a narrow waiver for certain kinds of safety conversations. He offers “checkpoints” as one possible approach: once a model has a given capability, it must come with certification of its alignment properties. For the third step, he lists levels of agreement in order of increasing difficulty, from banning clearly dangerous uses such as using AI to make biological weapons, to pre-release testing, to limiting the speed of recursive self-improvement, and finally to full pacing or even a pause. He supports floating that last level but thinks it is unlikely to happen any time soon.

Who Took a Public Position, and in What Form

Their own posts show that all three responded on X by referencing Amodei's announcement post, each giving their own view. They did not sign a shared document, so this should not be called a “co-signed statement,” and their positions differ in strength.

Altman's post is the most specific. He says he agrees that “we need to pace the frontier,” calling it one of the main topics OpenAI has discussed in recent weeks. He calls committing to give independent evaluators employee-like access “a great idea,” says OpenAI will do the same, and adds that there will be more to share soon. The post itself does not give a timeline, say which organization would do the evaluating, or define the scope of access.

Hassabis, for his part, says the piece points toward the right path forward: the details still need working through, but the direction is correct. He adds that this is also why he recently proposed an industry-wide standards body for frontier AI. Musk's response is a single line, “Dario is right,” with no mention of any measures. All three express agreement to varying degrees, but only Altman says his own company will follow the first step.

This article leaves out other people's responses, because their full original posts could not all be obtained. When you come across secondhand claims that someone else also supports the plan, it is worth clicking through to the original post, because “the direction is correct,” “we will do the same,” and “Dario is right” carry different levels of commitment.

Difficulties the Piece Acknowledges, and Questions Readers Can Ask

Amodei points out several limitations himself. Some forms of industry coordination are legally difficult and would need government support. He also believes limits on inputs such as training compute should be considered, but worries that some of these measures may be more “gameable” than a model's external behavior. He further writes that how far democracies can slow down is limited by the lead US companies hold over authoritarian regimes, chiefly the Chinese Communist Party, and argues for defending that gap through chip export controls, cracking down on unauthorized model distillation, and strengthening security to prevent the theft of model weights.

At the global level, he believes any agreement would need either extremely robust verification or a scope narrow enough that even if the other side broke it, it would not pose an existential military threat. The last two steps of the plan require industry coordination or government action and are not something a single company can decide, and the piece does not set any quantitative standard for how much to slow down.

The following questions were compiled by the editors and do not represent the position of any particular person: how the evaluating organizations will be chosen and how their independence will be ensured; how detailed the published findings will be and how much will be redacted; and how companies coordinating on standards will balance market competition with consumer interests. The piece discusses AI safety and technology policy toward China in the same text and argues that the two affect each other; readers can examine these two kinds of claims separately and decide for themselves whether they agree with each.

What It Actually Means for Everyday Users

Neither the piece nor the three responding posts mention shutting down existing services, removing features, or changing paid plans, and Amodei also writes that progress will still be relatively fast. Going by these original texts alone, nothing indicates an immediate change in how people currently use ChatGPT, Claude, or Gemini, and the piece itself does not say that any company will pause the launch of new models.

What bears more directly on general readers is transparency. If embedded evaluators really are brought in and publish their findings, there may in future be one more report to read, written by external reviewers and only narrowly redactable by the company. Amodei also writes that even though Anthropic's model cards and risk reports run to hundreds of pages, the company still decides what goes in and what is left out.

Here is an example designed by the editors. When a company compares AI tools internally, beyond features, pricing, and data terms, it could also add “whether external evaluation reports exist” and “whether incidents are publicly explained” to its comparison table. If older family members or students see a headline saying the AI industry is about to pause, you can go back to the source together and confirm: the piece argues for pacing capabilities advancement, not for halting model training or technical progress, and a “pause” is only the hardest level of global agreement to achieve.

2026 AI News Roundup: Highlights and Daily Life Applications from January to September

Claude Fable 5.1 and Mythos 5.1: Capabilities, Access, and Limits

• AI tools
• AI news