Dario Amodei, chief executive of Anthropic, the company behind the Claude family of AI models, published an essay on Saturday, September 12, 2026, calling for the AI industry to deliberately slow down how fast it improves model capabilities. The essay, titled “We Must Pace the Frontier,” lays out a three-step framework, and within a day Sam Altman, CEO of OpenAI, the maker of ChatGPT, signaled he was open to a version of the same idea. It is a rare moment: the two men running the two most-funded AI labs on the planet are, at least on paper, agreeing to hit the brakes.
The timing matters. For most of 2026, the story in AI has been about acceleration – bigger training runs, faster release cycles, and labs racing each other to ship the next frontier model before a rival does. Amodei’s essay reverses that framing. He argues that the industry should “slow the pace at which we improve the capabilities of AI models,” while being explicit that this is not a call to halt research or freeze training runs. It’s a call to pace them.
What Amodei’s Three-Step Plan Actually Proposes
The framework Amodei published on his personal site breaks the slowdown into three sequential steps, each harder to execute than the last. It is worth walking through them individually, because the difficulty curve is the whole story: step one is something Anthropic can do unilaterally tomorrow, and step three depends on governments that don’t trust each other agreeing on anything at all.
The first step calls for embedded evaluators – independent, third-party safety reviewers who get ongoing, employee-like access inside leading AI companies. These evaluators would sit close to the risk-assessment process, watch training pipelines, and have standing to flag and report incidents rather than reviewing a company’s safety work after the fact through a one-off audit. Anthropic has committed to this piece on its own, without waiting for competitors to sign on.
The second step asks frontier AI companies, particularly those based in democratic countries, to coordinate on shared safety standards and agree to limits on how fast unchecked capability growth is allowed to run. This is the step that requires competitors to trust each other, which is historically the hardest part of any industry self-regulation effort – antitrust concerns alone tend to keep rival CEOs from picking up the phone to coordinate on anything.
The third step is international: getting democratic governments to coordinate with authoritarian governments on managing AI risk globally. Amodei’s own framing acknowledges the difficulty here – verifying compliance across borders, especially with governments that have no obligation to be transparent about their AI programs, is a problem nobody has solved for any prior dual-use technology, nuclear included.
Amodei is careful to draw a line between pacing and stopping. “We must slow the pace at which we improve the capabilities of AI models,” he wrote, a framing that leaves plenty of room for continued investment, continued hiring, and continued product shipping – just not at the current rate of capability jumps between releases.
Sam Altman’s Response: Cautious Agreement, Not a Pact
Sam Altman’s reaction unfolded in two stages. According to reporting dated September 11, 2026, Altman told OpenAI staff at a company-wide meeting that the firm was open to slowing development of its most advanced systems, potentially alongside several other labs, though he acknowledged not everyone would agree to join. That’s a notably different posture than OpenAI took for most of the last two years, when the company’s public messaging leaned hard into shipping faster than rivals.
Then, after Amodei’s essay went public, Altman responded directly. On X, Altman wrote, “I agree with Dario that we need to pace the frontier.” The Atlantic’s coverage of the episode also noted Altman signaling support for bringing in independent evaluators, echoing the first and most concrete step of Amodei’s plan. Elon Musk, who runs xAI Corp., added his own two-word endorsement, “Dario is right,” pulling a third major lab into the conversation even though xAI wasn’t named in Amodei’s original framework.
What’s notable is what none of this amounts to yet. There’s no signed agreement, no joint statement, no binding commitment from OpenAI or xAI matching Anthropic’s unilateral move on embedded evaluators. What exists right now is a public essay from one CEO and public agreement, in principle, from two others. Whether that turns into a real coordination mechanism, or just becomes a talking point that fades by October, is the open question every AI safety researcher and every AI-dependent enterprise buyer is now watching.
Why This Is Happening Now
The proposal doesn’t land in a vacuum. It follows a stretch of high-profile incidents involving misuse of frontier models, and it lands on the heels of Anthropic’s own disclosures about Claude being implicated in cybersecurity incidents earlier this year – the kind of headline that puts a safety-first CEO in a stronger position to argue publicly for slowing down, since he’s not speaking from a position of having a clean record to defend.
There’s also a structural read on the timing. Frontier labs have spent 2026 locked in a release-cycle arms race: models arriving every few weeks, each one claiming a new state-of-the-art benchmark, each one requiring the next lab to respond within days to avoid looking like it’s falling behind. That cadence is exhausting for labs and confusing for enterprise buyers trying to plan around a moving target. A coordinated pacing framework, even a loose one, gives every CEO in the room a face-saving way to step off that treadmill without looking like they blinked first.
None of that makes the safety argument insincere. Amodei has been the most consistent voice among frontier lab CEOs on AI risk since Anthropic’s founding, and this essay is consistent with years of public writing on the subject. But the business logic and the safety logic point in the same direction here, which is probably part of why the proposal found quick, public agreement from two competitors instead of silence or pushback.
Historical Context: AI’s On-Again, Off-Again Relationship With Self-Restraint
This isn’t the industry’s first attempt at voluntary restraint, and the track record of past attempts is worth remembering before anyone assumes this one sticks. Open letters calling for AI pauses have circulated before, signed by researchers and executives, and generally produced headlines rather than changed release schedules. Voluntary safety commitments made at White House-brokered meetings in prior years likewise produced public pledges that were, in practice, difficult to verify and easy to quietly walk back once competitive pressure returned.
What makes Amodei’s proposal structurally different is the embedded-evaluator piece. A pledge to “be careful” is unenforceable. A commitment to give an outside team standing, ongoing access to your training pipeline is at least checkable – either the evaluators are in the building with real access, or they aren’t. That’s a lower bar than international treaty-style verification, but it’s a higher bar than prior industry pledges have cleared, and it’s the reason this proposal is getting more serious treatment from safety researchers than the open letters that came before it.
Market and Competitive Impact
For a story about slowing down, the market reaction has been notably calm. Neither Anthropic nor OpenAI is publicly traded, which removes the instant stock-price verdict that usually follows a major AI headline. But the proposal still has real competitive implications, because “pacing the frontier” is not a neutral policy for every lab in the market at the same rate.
A lab that is currently ahead on capability benefits more from everyone agreeing to slow down together than a lab that is behind and counting on a fast release cadence to catch up. That dynamic shapes how this proposal will likely be read by smaller and mid-tier labs – as well as by Chinese labs such as DeepSeek and Alibaba’s Qwen team, who were not named participants in Amodei’s framework and have no obvious incentive to slow their own release cadence just because two well-funded US labs have agreed to. If Anthropic and OpenAI pace their releases while other labs don’t, the practical effect could be a widening capability gap rather than an industry-wide slowdown – a dynamic that critics of past AI-pause proposals have raised before.
Enterprise buyers, meanwhile, may read this as a positive signal regardless of how it shakes out competitively. A slower, more predictable release cadence is easier to build a product roadmap around than the current environment, where a new frontier model can make a six-month-old integration look outdated overnight.
How the Major Labs Compare on Safety Posture
The table below summarizes each lab’s public position as of September 13, 2026, based on the reporting above. It reflects public statements only – none of this is independently verified or contractually binding at this stage.
| Lab / Leader | Public Position on Pacing | Concrete Commitment Made | Named in Amodei’s Framework |
|---|---|---|---|
| Anthropic (Dario Amodei) | Author of the “pace the frontier” proposal | Unilateral commitment to embedded, employee-like third-party evaluators | Yes – originator |
| OpenAI (Sam Altman) | Publicly agrees; told staff OpenAI is open to slowing its most advanced systems | Verbal support for independent evaluators; no signed commitment reported | Referenced in coverage, not the framework’s author |
| xAI (Elon Musk) | Publicly endorsed Amodei’s call (“Dario is right”) | None reported | Not named in the original three-step plan |
| DeepSeek, Qwen and other non-US labs | No public statement reported | None reported | Not named in the framework |
The Enforcement Problem
Every part of this plan gets harder to enforce as it scales up. Step one, embedded evaluators, is enforceable because it’s binary and local: either a third-party team has standing access inside a lab’s training process, or it doesn’t. Anthropic says it’s doing this now, which makes it checkable by outside observers relatively quickly.
Step two is much softer. “Coordinate on shared safety standards” is the kind of language that can mean anything from a formal joint body with real teeth to a handful of CEOs occasionally comparing notes at a conference. Without a named standards body, a named enforcement mechanism, or a named penalty for defection, step two is a statement of intent rather than a policy.
Step three, international coordination between democratic and authoritarian governments, is the hardest problem in the entire plan and Amodei’s own essay acknowledges as much. Nuclear non-proliferation took decades of treaty work, inspection regimes, and near-misses to get to its current (imperfect) state, and that was for a technology with far higher physical barriers to entry than a GPU cluster and a training script. There is no comparable inspection regime for AI capability, and no consensus yet on what one would even look like.
What Independent Evaluators Would Actually Do
The embedded-evaluator model borrows a structure familiar from other high-risk industries: financial auditors with standing access to a bank’s books, or safety inspectors with standing access to a nuclear plant’s operations, rather than periodic outside reviews. Applied to AI, that would mean a team that isn’t an employee of the lab but has employee-level visibility into training runs, risk assessments, and – critically – the ability to report an incident without needing the lab’s permission to do so first.
That last point is the one worth watching closely as this plays out. A reviewer who needs a lab’s sign-off before disclosing a problem isn’t independent in any meaningful sense; they’re a consultant. Whether Anthropic’s embedded evaluators get genuine, unilateral disclosure rights – and whether OpenAI adopts an equivalent structure rather than a weaker version – will tell you more about how serious this proposal is than any public statement from either CEO.
Reactions From the Safety Research Community
Amodei’s essay lands at a moment when independent evaluator groups – organizations that already do third-party model testing on a contract basis – are positioned to be the natural candidates for the embedded-evaluator role his plan describes. The framework doesn’t name specific organizations publicly beyond describing the kind of access it envisions, but the model closely resembles existing third-party red-teaming arrangements, just with standing access instead of periodic engagements.
For researchers who have spent years arguing that self-regulation without verification is close to meaningless, the appeal of this plan is specific: it’s the first major CEO-level proposal that ties a public safety pledge to a concrete, checkable mechanism rather than a vague promise to “be careful.” That’s a meaningfully different ask than prior open letters, even if steps two and three remain aspirational for now.
Expert Perspectives
Dario Amodei, CEO of Anthropic, framed the core problem directly in his essay: “We must slow the pace at which we improve the capabilities of AI models.” (source)
He laid out the goal behind the plan in more detail: “I’m therefore proposing a three-step plan with the goal of pacing the frontier: building AI at a balanced rate that aims to ensure its safety while still achieving its benefits and grappling with important geopolitical dilemmas.” (source)
On the first, concrete step of the plan, the BBC reported that Amodei said Anthropic would commit to bringing in outside monitors immediately, without waiting on rivals: “The first step in his plan entails AI labs committing to bringing in outside monitors to look out for hazards; he said Anthropic will commit to that right now and hopes rivals will join.” (source)
Sam Altman, CEO of OpenAI, responded on X with direct agreement: “I agree with Dario that we need to pace the frontier.” (source)
The Atlantic reported that Altman went further, backing the specific mechanism Amodei proposed: “Committing to having independent evaluators.” (source)
Timeline of Public Statements
| Date | Event | Source |
|---|---|---|
| Week of September 7, 2026 | Sam Altman tells OpenAI staff at a company-wide meeting the firm is open to slowing development of its most advanced systems | Reported September 11, 2026 |
| September 12, 2026 | Dario Amodei publishes “We Must Pace the Frontier,” proposing the three-step framework | darioamodei.com |
| September 12, 2026 | Elon Musk publicly backs the proposal on X | Los Angeles Times |
| September 12, 2026 | Sam Altman publicly agrees, endorses independent evaluators | The Atlantic |
| September 13, 2026 | Global outlets report on the three-way public alignment | Moneycontrol |
What Enterprise AI Buyers Should Watch For
For companies building on top of Claude, ChatGPT, or Grok, this story is less about ideology and more about roadmap risk. A genuine industry-wide pacing agreement, if it holds, would mean fewer surprise capability jumps and more predictable release cycles, which is generally good news for teams that have to plan integration and testing work around model updates.
The flip side is competitive risk if the pacing agreement is asymmetric. If Anthropic and OpenAI pace their releases while labs outside this framework, including major Chinese labs, don’t, buyers who standardized on a “safety-first” US lab could find themselves behind on raw capability compared to competitors using a faster-moving alternative. That’s a real strategic tradeoff, not just a safety-versus-speed talking point, and it’s one procurement teams evaluating AI vendors should factor into multi-vendor strategies rather than assuming today’s leaderboard position holds steady.
Predictions: Where This Goes From Here
- Anthropic will publicize details of its embedded-evaluator program within weeks, likely naming the scope of access granted, since a unilateral commitment made in a public essay carries reputational pressure to follow through visibly.
- OpenAI’s actual commitment will likely lag its rhetoric. Expect a formal policy statement or blog post from OpenAI within the next one to two months, but expect it to be less specific on access rights than Anthropic’s commitment.
- Step two – coordination among frontier labs on shared standards – will move slowly and probably won’t produce a binding joint agreement in 2026. Antitrust sensitivities and competitive distrust make fast multilateral coordination between rival CEOs unlikely on this timeframe.
- Labs outside this framework, including major Chinese AI developers, will not adopt equivalent pacing commitments in the near term, since they were not named participants and have separate competitive incentives to keep shipping at their current pace.
- Expect renewed scrutiny of what “pacing” means in practice the next time any of these three labs – Anthropic, OpenAI, or xAI – ships a major capability jump, since critics will measure the rhetoric of September 2026 against the release cadence that follows it.
The Skeptical Read
It’s worth stating the obvious counter-argument plainly: public agreement costs nothing. Altman’s post on X took seconds to write and commits OpenAI to nothing enforceable. Musk’s two-word endorsement is even lighter. The only party that has made a concrete, checkable commitment so far is Anthropic, and Anthropic is also the party whose CEO wrote the essay – meaning the one binding action taken this week was also the one most obviously in service of building goodwill for the company that took it.
None of that means the proposal is cynical. Amodei’s public record on AI safety predates this essay by years, and the embedded-evaluator mechanism is a genuinely more verifiable structure than past industry pledges. But readers should treat “pacing the frontier” as a framework under construction, not a settled policy, until there’s a second lab with a comparably concrete, independently checkable commitment to point to.
Frequently Asked Questions
What did Dario Amodei actually propose?
A three-step framework, published September 12, 2026, calling for embedded independent safety evaluators inside AI labs, coordination among frontier companies on shared safety standards, and international cooperation between governments to manage AI risk.
Did Sam Altman agree to slow down OpenAI’s AI development?
Altman told OpenAI staff the company was open to slowing development of its most advanced systems, and publicly wrote on X that he agrees with pacing the frontier. He has not announced a formal, binding policy matching Anthropic’s evaluator commitment.
Is this an official agreement between OpenAI and Anthropic?
No. There is no signed, joint agreement between the two companies. Anthropic has made a unilateral commitment; OpenAI’s CEO has expressed public support.
Does “pacing the frontier” mean AI companies will stop releasing new models?
No. Amodei explicitly said pacing is not the same as halting training or technical progress – the goal is a slower rate of capability improvement, not a freeze.
What is an “embedded evaluator” in this context?
A third-party safety reviewer given ongoing, employee-like access inside an AI lab to monitor training and risk-assessment processes and report incidents independently, rather than conducting periodic outside audits.
Is Elon Musk’s xAI part of this framework?
Musk publicly endorsed Amodei’s call for a slowdown, but xAI was not named as a participant in the original three-step plan and has not announced a concrete commitment.
Will Chinese AI labs like DeepSeek join this pacing effort?
There is no public indication they will. The framework targets coordination among labs in democratic countries as its second step, and no non-US lab has been named as a participant so far.
Why does this matter for businesses using AI models?
A more predictable, paced release cycle from major labs could make it easier to plan product roadmaps around model capabilities, though buyers should weigh the risk of falling behind if competitors adopt faster-moving, non-participating labs instead.
