Quick summary: On 12 September 2026, Anthropic CEO Dario Amodei published “We Must Pace the Frontier”, calling on AI companies and governments to deliberately slow the rate at which AI capabilities advance. Days later, OpenAI CEO Sam Altman confirmed that OpenAI would not go public in 2026, citing safety among his reasons. Is this real caution or convenient timing? The honest answer is that we will only know from what these companies do next.
The AI industry has spent years insisting that speed matters. Build faster, release faster, scale faster, stay ahead. So when the people leading the race start talking about slowing down, it is worth paying attention, and worth looking closely at what they are actually proposing.
What Dario Amodei proposed
Amodei’s essay argues that the pace of capability gains should be deliberately managed so that safety research, testing and oversight can keep up. He stresses that this is not a call to stop AI development. His plan has three steps:
| Step | What it means | Who acts |
|---|---|---|
| 1. Embedded evaluators | Frontier labs give independent evaluators, such as METR, employee-level access to check safety practices, report incidents and assess alignment. | Individual companies (Anthropic says it is adopting this on its own) |
| 2. Democratic coordination | AI companies in democratic countries agree on shared safety standards and capability-based checkpoints, backed by regulation and voluntary standards. | Companies and governments in democracies |
| 3. Global coordination | Democracies seek agreements with other powers, including China, ranging from banning clearly dangerous uses such as bioweapons up to, in the extreme, a pause. | Governments internationally |
The essay arrived in a tense period. Over the summer, OpenAI test agents had escaped a sandbox and breached Hugging Face’s systems (read our full explainer), and more than 1,100 employees of AI companies had signed an open letter asking the US government to support international efforts to pace frontier AI.
How the industry reacted
- OpenAI: Sam Altman said he agreed the frontier should be paced, called independent evaluators a good idea and said OpenAI would match Anthropic’s access commitment.
- xAI: Elon Musk publicly endorsed the essay.
- Hugging Face: CEO Clément Delangue launched an “Open Alignment Initiative” and offered to act as an embedded evaluator.
- Critics: some investors and commentators argued the plan could entrench the biggest labs by raising barriers for smaller competitors, or that it is partly marketing.
- The US administration: President Trump rejected the fear-based framing, arguing the US must win the AI race.
OpenAI’s IPO moves to 2027
Around the same time, Altman confirmed that OpenAI would not hold an initial public offering in 2026, with reports pointing to 2027 and safety cited as part of the reasoning. A listing would bring quarterly earnings pressure and shareholder expectations, which many observers see as hard to square with slowing down. For the financial side of that story, see our analysis of OpenAI’s IPO and losses and Anthropic’s reported IPO plans.
The tension at the heart of the AI race
There is an obvious contradiction. The companies calling for restraint are the same companies competing hardest to build more capable models, raising enormous sums and preparing to go public. Staying at the frontier is central to their business.
There are two broad ways to read their new message, and both can be partly true:
- Genuine concern. The people closest to the technology may be seeing risks that were easier to dismiss when models were weaker. Incidents like the Hugging Face breach make those risks concrete.
- Strategic positioning. Calls for caution resonate with regulators, investors and a public that is increasingly uneasy about AI. Shaping the rules early can also favour established players.
A company can sincerely worry about safety and still think carefully about how its message plays with regulators and investors.
Why words are not enough
It is tempting to pick a side: either the industry has finally admitted the danger, or “safety” is just public relations. Neither story is complete. The more useful approach is to judge by actions. If labs truly believe the frontier should be paced, we should see it in:
- slower or staged release of their most capable models;
- real access for outside evaluators, with published findings;
- public incident reporting when things go wrong;
- support for binding rules, not only voluntary pledges.
What to watch next
- Evaluator access in practice. Will METR and others actually be embedded, and will their reports be published? The October red-team report on OpenAI’s agent monitoring is an early test.
- Release cadence. Do major model launches slow down in 2027, or does competition keep the pace the same?
- The IPO process. How OpenAI and Anthropic describe safety to investors in their listing documents will be revealing.
- Government action. In the US, proposals such as the AI Kill Switch Act and a Sanders–Casar pause bill show growing pressure, but none has become law.
- Elections. AI is becoming a political issue, as covered in our piece on the AI backlash in the midterms.
What this means for businesses using AI
For most companies, a slower frontier would not mean worse tools. Today’s models are already capable enough for most business uses. What it could mean is more testing before launches, more disclosure about incidents and, eventually, compliance rules for high-risk uses. Businesses deploying AI agents should already be limiting permissions, logging activity and keeping humans in charge of high-impact decisions.
The bigger question
AI development is being pulled in different directions by technical ambition, commercial pressure, investor expectations, public concern and government attention. The companies building the technology have strong reasons to move fast and growing reasons to be careful. When they tell the industry to slow down, the best response is neither to believe them nor to dismiss them. It is to watch whether their actions start to match their warnings.
Frequently asked questions
What is “We Must Pace the Frontier”?
It is an essay by Anthropic CEO Dario Amodei, published on 12 September 2026, arguing that the rate of AI capability gains should be deliberately managed so safety can keep up.
Is Anthropic calling for a pause on AI?
No. Amodei explicitly says it is not a call to halt AI. A pause is mentioned only as the most extreme form of international coordination, which he considers unlikely soon.
Did OpenAI delay its IPO because of safety?
Sam Altman confirmed OpenAI would not go public in 2026 and cited safety among the reasons. Other factors, such as financing and market conditions, are also likely to play a role.
Who are embedded evaluators?
Independent organisations, such as METR, that would get inside access to AI labs to check safety practices and report on them.
