Dario Amodei spent Saturday telling the industry he built to slow down. The Anthropic CEO published a roughly 3,800-word essay titled “We Must Pace the Frontier” on his personal site on September 12, 2026, arguing that AI labs need to deliberately throttle how fast they improve model capabilities so safety research has time to catch up. Within hours, OpenAI CEO Sam Altman posted his agreement on X. By September 13, Chosun Ilbo was running the story as “OpenAI, Anthropic CEOs Propose AI Development Brake,” and outlets from Reuters to the BBC to Stocktwits had picked up the thread. This is not a repeat of the vague pause letters of 2023. It is the first time two sitting frontier-lab CEOs have put a structured, three-step operational plan on the table, and markets, rival labs, and Washington are all reacting differently to it.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Amodei Actually Proposed in “We Must Pace the Frontier”
Amodei’s essay, published on darioamodei.com on September 12, 2026, opens with a blunt thesis. “We must slow the pace at which we improve the capabilities of AI models,” he wrote. He was careful to distinguish this from the 2023-era calls for an outright freeze on training runs. Pacing, in his framing, means slowing the rate of capability jumps, not stopping technical work altogether. “Progress will still seem fast, and we must make wise use of the time we gain,” he added in the same post.
The essay reportedly warns that swarms of AI agents, if left insufficiently constrained, could cause internet-scale damage within six to twelve months, according to Axios’s September 12 report on the piece. Amodei’s stated goal is to buy the industry an additional one to two years to build out alignment and interpretability tooling before frontier models cross into riskier territory. That timeline, more than any single number, is what separates this proposal from the symbolic gestures AI executives have made in years past.
Altman’s Endorsement Landed Within Hours
Sam Altman did not wait long to respond. On the same day Amodei’s essay went live, Altman wrote on X: “I agree with Dario that we need to pace the frontier.” He went further than a simple nod of approval, telling the two labs’ shared safety proposal was actionable now. “Committing to having independent evaluators with employee-like access is a great idea, and we will do the same,” Altman wrote, according to a report from The Atlantic.
Altman’s alignment with Amodei is not entirely new. Back in a July 28, 2026 appearance on the Invest Like the Best podcast, he had already floated similar language, telling TechCrunch: “We may have to pace the rate of AI development to give ourselves enough time for society to harden around some of these new capability levels.” Reuters had also reported on September 10 and 11 that Altman told OpenAI staff internally that the company was open to slowing its own development timeline, citing Bloomberg’s sourcing. Taken together, the sequence suggests Altman’s public endorsement of Amodei’s essay was less a sudden reversal than the culmination of weeks of internal deliberation at OpenAI.
Inside the Three-Step Slowdown Framework
What makes Amodei’s essay different from the open letters of the past is structure. He laid out three concrete steps rather than a single demand, and he made clear that only the first one is something Anthropic can execute unilaterally right now.
Step One: Independent Evaluators With Employee-Level Access
The first step calls for embedding third-party evaluators, organizations of the kind that already audit frontier labs on a contract basis, with permanent, employee-like access inside the companies building the most powerful models. That means desks, credentials, and technical permissions similar to what an internal risk team would get, plus a contractual right to publish findings on incidents without the lab editing the results first. Amodei says Anthropic is moving on this step alone, without waiting for the rest of the industry. Altman’s public commitment to match it means OpenAI is now on the hook for the same arrangement, at least rhetorically.
Step Two: Shared Safety Standards Among Democratic-Country Labs
The second step is harder, because it requires competitors to coordinate. Amodei wants major labs based in democratic countries to agree on shared safety checkpoints and speed limits for capability gains, likely with government backing. Because direct coordination between competing companies on product roadmaps can trigger antitrust scrutiny, the essay explicitly calls for narrow antitrust waivers or clear government support so labs can discuss safety-related limits without legal risk. This is the piece of the plan that has already drawn political pushback, including from figures inside the current administration who argue existing product-liability law and market competition already discipline the labs without new carve-outs.
Step Three: Global Coordination, Including China
The third and most ambitious step reaches beyond democratic-country labs to propose limited coordination with China on the narrowest, highest-stakes risks, such as preventing AI-assisted development of biological weapons and testing frontier models for acute cybersecurity and alignment risks before public release. The scale of what Amodei is asking Beijing to throttle is not hypothetical: Alibaba pushed its Qwen3.8-Max model into public availability in August 2026 packing 2.4 trillion total parameters, with 95 billion active at any given time, a frontier-scale release that shows little appetite among Chinese labs for voluntary deceleration. Amodei reportedly acknowledges in the essay that full cooperation with Beijing is unlikely, but argues that even partial, verifiable agreement on the worst-case scenarios is worth pursuing.
How Rival Labs and Executives Reacted
Reaction from the rest of the industry came fast, if not uniformly deep. Elon Musk posted a two-word reply on X, “Dario is right,” a response multiple outlets picked up as a sign that xAI’s leadership was not going to fight the pacing framework publicly, even though Musk’s own company has built its reputation on rapid iteration. Google DeepMind CEO Demis Hassabis reportedly also voiced general support for the idea of pacing frontier development, joining Amodei, Altman, and Musk in what one outlet framed as a rare moment of public agreement among AI’s biggest rivals.
Inside OpenAI, the position is not coming only from Altman. Chief scientist Jakub Pachocki has separately argued that AI companies should be coordinating to slow down future development as needed, calling for voluntary slowdowns until the industry can agree on shared safety standards, a position Reuters reported on September 9. Anthropic co-founder Jack Clark had previewed the underlying metaphor months earlier, telling BBC Newsnight in June that the AI industry has “a gas pedal, but it doesn’t have a brake pedal,” a line that anticipated the “development brake” framing used by Chosun Ilbo and other outlets this week.
| Lab / Figure | Role | Public Stance on Pacing | Commitment Level |
|---|---|---|---|
| Anthropic (Dario Amodei) | CEO | Authored the three-step framework; wants slower capability gains industry-wide | Unilateral: embedding independent evaluators now |
| OpenAI (Sam Altman) | CEO | Publicly endorsed the essay and the evaluator proposal | Matching commitment on evaluators, pledged |
| OpenAI (Jakub Pachocki) | Chief Scientist | Backs voluntary, coordinated slowdowns among labs | Advisory position, no formal pledge reported |
| xAI (Elon Musk) | CEO | Brief public endorsement of Amodei’s essay | No specific commitment reported |
| Google DeepMind (Demis Hassabis) | CEO | Reported general support for pacing frontier development | No specific commitment reported |
| Anthropic (Jack Clark) | Co-founder | Originated the “brake pedal” framing in June 2026 | Conceptual, predates formal proposal |
Why Amodei Is Sounding the Alarm Now
The timing is not incidental. Both Anthropic and OpenAI are in the middle of the largest capital raises in their history, against a backdrop Stanford HAI’s 2026 AI Index pegs at $285.9 billion in U.S. private AI investment for 2025 alone, and both were widely expected to be closing in on public listings this year. Anthropic closed a $65 billion Series H round at a $965 billion post-money valuation in May 2026, according to TechCrunch, and confidentially filed for a U.S. IPO on June 1, 2026, per Reuters. By September 11, Reuters reported Anthropic was seeking to raise as much as $100 billion at a valuation approaching $2 trillion, with Nvidia reportedly weighing an investment of up to $10 billion as part of that round. Anthropic has kept shipping through the fundraising blitz rather than pausing to match its own rhetoric: the company introduced Fable 5.1 in September 2026 with a 75% cut to cache-read pricing and a reported 52.6% score on the Terminal-Bench-Science benchmark, according to OriginBrief, evidence that the slowdown essay was never meant to apply to Anthropic’s own release schedule.
OpenAI’s numbers tell a similar story. The company closed a $122 billion funding round at an $852 billion post-money valuation on March 31, 2026, and later filed confidentially for its own IPO, reportedly targeting a valuation near $1 trillion. Some reporting suggests OpenAI’s finance chief has leaned toward pushing the actual listing into 2027 rather than debuting this fall. The same month Amodei’s essay circulated, OpenAI was pulling a very different lever: after CNBC reported a roughly 50% API price cut across its Sol and Luna line, the company launched its successor GPT-6.1 Sol model in September 2026 priced at about $2 per million input tokens and $10 per million output tokens, according to Capital & Compute, carrying that discount straight through into the new release. Publishing a safety-first slowdown essay while simultaneously racing to undercut rivals on cost, and while both companies are courting public-market investors and sovereign-scale valuations, is a signal to regulators and the public as much as it is a technical proposal.
How Markets Responded to the Slowdown Call
The market reaction has been real but contained. Prediction-market contracts tied to OpenAI and Anthropic reportedly slipped by roughly 7% and 2.8% respectively in the hours after the essay circulated, though that movement was concentrated in speculative betting markets rather than public equities. Nvidia’s stock ended the week near $218, down from a month-to-date high of $234, a move some analysts tied partly to the slowdown headlines weighing on chip-sector sentiment. Big-tech names told a more mixed story the same day: Microsoft rose roughly 0.65%, Amazon climbed about 1.94%, Alphabet gained close to 1.53%, while Oracle slipped around 1.74%, according to Yahoo Finance’s September 12 coverage.
Portfolio managers quoted in market coverage described the essay as a short-term sentiment drag on chip and memory names rather than evidence of a broken AI trade. Capital expenditure plans from Microsoft, Amazon, Alphabet, and Oracle on AI infrastructure have not visibly slowed, and the consensus view reported across outlets is that the long-cycle AI buildout remains intact even as individual stocks wobble on headline risk.
| Metric | Value | Timeframe / Source |
|---|---|---|
| Nvidia share price | ~$218, down from a monthly high of ~$234 | Week of Sept 8-13, 2026 |
| Microsoft share move | +0.65% | Sept 12, 2026 |
| Amazon share move | +1.94% | Sept 12, 2026 |
| Alphabet share move | +1.53% | Sept 12, 2026 |
| Oracle share move | -1.74% | Sept 12, 2026 |
| Anthropic valuation | ~$965 billion post-money, targeting near $2 trillion at IPO | May 2026 raise; Sept 11, 2026 IPO reporting |
| OpenAI valuation | $852 billion post-money, targeting near $1 trillion at IPO | March 31, 2026 raise |
Echoes of 2023: How This Differs From the FLI Pause Letter
This is not the AI industry’s first brush with a public slowdown pitch. In March 2023, the Future of Life Institute’s open letter called for a six-month pause on training systems more capable than GPT-4, gathering thousands of signatures including many outside the frontier labs themselves. A few months later, the Center for AI Safety published a one-sentence statement, warning that mitigating the risk of extinction from AI should be a global priority alongside pandemics and nuclear war. Sam Altman, Dario Amodei, and Demis Hassabis were all among the signatories of that 2023 statement, meaning today’s proposal is the same cast of executives returning to the same concern three years later.
The difference this time is specificity. The 2023 letters asked for a blanket pause with no operational mechanism to enforce it, and no lab actually stopped training. Amodei’s 2026 framework rejects a pause outright and instead proposes a standing structure: embedded evaluators, cross-lab safety checkpoints, and narrow international coordination on the highest-risk capabilities. It reads less like a protest and more like a policy draft, which is part of why it is being treated as a bigger story than the symbolic gestures that preceded it.
The Political Pushback Has Already Started
Not everyone in Washington is on board with the antitrust-waiver piece of the plan. Within a day of Amodei’s essay circulating, prominent voices close to the administration’s AI policy apparatus pushed back publicly, arguing labs should pace themselves without asking for special legal carve-outs from the government. That pushback focused narrowly on step two of the framework, the part that requires labs to coordinate directly, rather than on the underlying idea of slowing capability gains, which figures across the political spectrum have been more willing to endorse in principle.
What Critics Say About Voluntary Pacing
Skeptics of the plan raise a familiar problem: voluntary commitments from companies racing each other for market share and investor capital are hard to enforce and easy to walk back once competitive pressure builds. Anthropic and OpenAI are both mid-IPO, competing for the same enterprise customers, and both need to keep shipping more capable models to justify trillion-dollar valuations to public investors. Critics argue that a pledge to slow down capability gains is difficult to square with the need to keep releasing headline-grabbing upgrades on a regular cadence to sustain investor enthusiasm.
There is also the China question. Step three of Amodei’s plan depends on cooperation from a government that has no obligation to slow its own AI programs just because two American CEOs asked nicely, and the pace of Chinese releases makes the ask harder with each passing month: DeepSeek shipped its V4 Flash 0731 model under an MIT license in July 2026 priced at just $0.14 per million tokens, a release built for maximum adoption rather than restraint. Even Amodei’s own essay reportedly concedes that full cooperation with Beijing is unlikely, which leaves the most ambitious part of the framework as more of an aspiration than a plan with a clear path to execution.
What This Means for Developers and Enterprise AI Buyers
For engineering teams building on top of OpenAI or Anthropic’s APIs, the practical near-term effect is likely to be modest, and if anything the opposite of a slowdown is showing up in pricing across the whole sector. Amodei has been explicit that pacing does not mean halting model releases, so the cadence of new API-accessible models is unlikely to freeze overnight, a point underscored by OpenAI’s September 2026 launch of GPT-6.1 Sol at roughly $2/$10 per million input/output tokens, Google’s September 2026 release of Gemini 3.8 Flash at $0.75/$3.75 per million tokens, and Anthropic’s own October 2026 launch of Claude Haiku 5.5 at a mere $0.10/$0.50 per million tokens, according to pricing tracked by Digital Applied and Capital & Compute. What could change is the depth of pre-release testing and the amount of red-teaming disclosure that ships alongside new frontier models, since the embedded-evaluator proposal is the one piece of the plan both companies have already agreed to implement. Enterprise buyers evaluating AI vendors for regulated industries may find that additional third-party audit trail attractive, since it offers an external verification layer that internal safety teams alone cannot provide, even as falling per-token costs make it cheaper than ever to run those models at scale.
Predictions: Where This Goes From Here
Several outcomes look more likely than others based on how the framework is structured and how the industry has responded to similar proposals in the past.
- Step one gets implemented first and fastest. Because Anthropic is already moving unilaterally and Altman has publicly matched the commitment, expect concrete announcements about specific third-party evaluator partnerships within the next two to three months.
- Step two stalls on antitrust concerns. Without an explicit government waiver, expect labs to describe shared “principles” rather than binding, coordinated speed limits, since direct coordination on product timelines carries real legal risk for competing companies.
- Step three remains aspirational through 2026. Meaningful coordination with Chinese AI labs on release-testing standards is unlikely to materialize this year, even if diplomatic conversations begin.
- IPO timing gets more cautious language, not delayed filings. Both companies have too much invested in their public-listing plans to slow them meaningfully, but expect more safety-focused messaging in roadshow materials as a hedge against public scrutiny.
- Other labs face pressure to respond. With Google DeepMind and xAI’s leaders already voicing support in principle, expect Meta and other frontier players to face growing questions about whether they will adopt similar third-party evaluator arrangements.
Competitive Landscape: Who Has the Most to Lose or Gain
Anthropic arguably has the most credibility to gain from moving first on independent evaluators, since safety positioning has long been part of its brand differentiation against OpenAI. For OpenAI, matching the commitment publicly costs little in the near term but sets a precedent the company will be expected to honor as it approaches its own public listing. xAI and Google DeepMind, by contrast, have offered supportive words without matching operational commitments so far, which leaves them with more flexibility but also more exposure if critics start asking why they have not embedded evaluators of their own. Meta, which was not reported as part of the public endorsements this week, faces a similar question given its aggressive open-weight model strategy, which sits somewhat apart from the frontier-lab framing Amodei’s essay addresses. Mistral is a reminder that the competitive pressure is not confined to the three largest U.S. labs: the French startup pushed its Large 4 model into public preview in October 2026 with roughly 1 trillion parameters, priced at just $0.68 per million input tokens and $2.09 per million output tokens according to The AI Rankings and Capital & Compute, aggressive enough to keep undercutting the very labs now debating whether to slow down.
Historical Context: A Pattern of Escalating Warnings
Industry warnings about AI risk have escalated steadily since 2023, moving from broad, signature-gathering letters to increasingly specific proposals from people who actually run the labs in question. The 2023 FLI letter and CAIS statement were followed by years of incremental safety announcements, congressional testimony, and voluntary commitments brokered by the White House. What sets September 2026 apart is that the proposal came directly from a sitting CEO’s personal essay rather than a third-party advocacy campaign, and that it was met with same-day, on-the-record agreement from a direct competitor. That kind of public alignment between Anthropic and OpenAI, companies that compete aggressively for the same enterprise contracts and the same pool of AI researchers, is rare enough on its own to explain why the story spread as fast as it did.
For more on the political fight over this exact proposal, see our coverage of David Sacks rejecting the AI slowdown deal. On the money side, read about Anthropic’s IPO plans and internal AI risk estimates, Mistral’s âŽ21 billion valuation after its Samsung deal, and an OpenAI scientist’s warning on AI risk days after the GPT-6 Astra launch. For background on OpenAI’s other 2026 controversies, see our report on the RubyGems rogue AI agent attack, and for OpenAI’s newest model, how to try GPT-6 Astra. Browse more coverage in our AI models hub.
Frequently Asked Questions
What is the “AI development brake” that OpenAI and Anthropic’s CEOs are proposing?
It refers to a three-step framework proposed by Anthropic CEO Dario Amodei in his September 12, 2026 essay “We Must Pace the Frontier,” which calls for deliberately slowing the rate of AI capability gains through independent model evaluators, shared industry safety standards, and limited international coordination, rather than halting AI development outright.
Did Sam Altman actually agree to slow down OpenAI’s AI development?
Altman publicly endorsed Amodei’s essay on X, writing that he agreed on the need to “pace the frontier” and committing OpenAI to matching Anthropic’s plan to give independent evaluators employee-level access. He has not announced a halt or delay to any specific OpenAI model release.
Is this the same as the 2023 AI pause letter?
No. The 2023 Future of Life Institute letter called for a temporary, blanket pause on training systems more powerful than GPT-4. The 2026 proposal explicitly rejects a pause and instead proposes an ongoing structure of embedded evaluators and coordinated safety checkpoints while development continues.
How did the stock market react to the AI slowdown announcement?
The reaction was mixed and modest. Nvidia’s share price dipped from a monthly high near $234 to around $218 during the week, while Microsoft, Amazon, and Alphabet posted small gains and Oracle slipped slightly on September 12, 2026. Analysts described the move as sentiment-driven rather than a sign of a broken AI investment cycle.
What is “pacing the frontier” supposed to prevent?
According to reporting on Amodei’s essay, the concern is that insufficiently constrained AI agents could cause significant, internet-scale disruption within six to twelve months if safety and alignment research does not keep pace with capability gains. Slowing the rate of improvement is framed as a way to buy roughly one to two additional years for protective measures to mature.
Are other AI companies like Google, Meta, or xAI joining the pledge?
Elon Musk and Google DeepMind CEO Demis Hassabis have both voiced general support for the idea of pacing frontier development, but as of this reporting neither company has announced the specific operational commitments that Anthropic and OpenAI have, such as embedding independent evaluators with employee-level access.
Does this affect OpenAI’s and Anthropic’s IPO plans?
Not directly, based on current reporting. Anthropic has been reported seeking to raise as much as $100 billion at a valuation approaching $2 trillion, and OpenAI has targeted a valuation near $1 trillion in its own IPO process, part of a U.S. AI funding environment Stanford HAI’s 2026 AI Index measured at $285.9 billion in private investment for 2025 alone. Neither company has announced a delay tied specifically to the slowdown proposal, though some reporting suggests OpenAI’s listing timeline may already be leaning toward 2027.
Who are the independent evaluators that would get access to AI labs?
Amodei’s proposal describes third-party organizations similar to existing AI model auditors gaining permanent, employee-like access, including credentials and technical permissions, along with a contractual right to publish safety findings without the lab controlling the final report.
