Four leaders at the frontier of artificial intelligence backed slowing the pace of model development on September 12, 2026. Anthropic CEO Dario Amodei made the detailed case. Within nine hours, SpaceXAI chief Elon Musk, OpenAI CEO Sam Altman and Google DeepMind chair Demis Hassabis had publicly supported its direction. That convergence is new. It is also narrower than a binding AI slowdown: only Anthropic made a specific unilateral commitment, and none of the four announced a shared schedule, capability ceiling or enforcement system.
Anthropic
OpenAI
xAI
GoogleWhat the four AI leaders actually agreed to
Amodei's essay, titled "We Must Pace the Frontier," argued that safety work needs time to catch up with capabilities. He proposed three layers: permanent third-party evaluators embedded inside frontier labs, coordination among companies in democracies, and eventually coordination with geopolitical rivals. Anthropic committed to the first step, giving outside evaluators ongoing, employee-like access to assess training processes, verify safety practices and report incidents.
The other endorsements varied. Musk wrote that Amodei was right. Altman agreed that the frontier should be paced and said OpenAI would also give external evaluators access. Hassabis said the essay pointed in the right direction while noting that the details still needed work. Those statements create unusual agreement on the problem and broad direction, but not yet on the operational rules.
Nine hours from proposal to four aligned statements
Reported sequence on September 12, 2026. Times are US Eastern Time.
Method: elapsed time is measured from Amodei's 10am publication. Source: Axios reporting on the four public statements, September 13, 2026.
Why this is not the July employee letter
The July "Pacing the Frontier" letter was an employee petition that had 1,178 signatures when we covered it and now lists 1,386. It asked the US government to support international work on tools for deliberately pacing automated AI development. The September event moves the argument from employees to lab leadership and adds a concrete institutional mechanism: embedded evaluators with access to internal processes.
It still stops short of a moratorium. Amodei explicitly said pacing does not mean ending model training or technical progress. His preferred model is conditional: when a system reaches a defined capability, such as reliably defeating common sandboxes, it would need to demonstrate corresponding safety properties before advancing further.
Embedded evaluators are the testable part
The strongest element of the proposal is not the word "slow." It is the attempt to make voluntary restraint observable. An evaluator with continuing access to model training, internal tests and incident records could compare a lab's public policy with what happens inside its development pipeline. That is more informative than testing only a finished model immediately before release.
There are unresolved questions. Who selects and pays the evaluator? What can be published when findings involve security-sensitive systems? Can an evaluator stop a training run, or only report concerns? How would access work without exposing model weights or personal data? Until those powers and boundaries are defined, "employee-like access" describes proximity, not authority.
The agreement is meaningful because four competitors named the same direction. It is incomplete because they did not name the same brake.
On what changed on September 12
The coordination problem remains
Any voluntary pacing system faces a familiar incentive problem. A lab that slows alone may lose researchers, customers and strategic ground to competitors that do not. Amodei acknowledged that democratic countries cannot slow beyond their lead over authoritarian rivals without creating a national-security risk. He therefore argued that government may need to enable safety coordination, including addressing antitrust barriers, and eventually negotiate internationally.
The political response showed why implementation will be difficult. President Donald Trump said on September 13 that some guardrails may be appropriate but resisted slowing US development if it allowed China to catch up. Congressional leaders discussed bringing AI executives and lawmakers together, but no binding framework emerged over the weekend.
This leaves a clear way to judge what comes next. Look for named evaluators, written access terms, common capability thresholds, public incident-reporting rules and a government mechanism that allows competitors to coordinate on safety without coordinating commercially. If those appear, September 12 may mark the start of a real governance system. If they do not, the nine-hour sequence will remain an important statement of concern rather than an operational slowdown.
Sources:
- We Must Pace the Frontier (Dario Amodei, Sept 2026)
- Anthropic, OpenAI CEOs call for slowdown in AI development (Axios, Sept 12, 2026)
- Hitting the AI brakes: four lab leaders endorse pacing (Axios, Sept 13, 2026)
- Anthropic CEO says AI safety measures need time to catch up (AP, Sept 12, 2026)
- Trump downplays slowing AI development amid China competition (AP, Sept 13, 2026)
- House Speaker calls for an AI meeting but says Congress will not lead (Axios, Sept 13, 2026)
- Pacing the Frontier employee statement
Image credits
- Hero: Dario Amodei speaking at TechCrunch Disrupt in San Francisco on September 20, 2023. Photo by Kimberly White for TechCrunch, Wikimedia Commons, licensed CC BY 2.0.
- In-body: Rishi Sunak meeting Dario Amodei, Demis Hassabis and Sam Altman at 10 Downing Street on May 24, 2023. Photo by Simon Walker / No 10 Downing Street, Wikimedia Commons, licensed CC BY 2.0.