Amodei’s ‘We Must Pace the Frontier’ essay proposes a specific, three-stage AI safety architecture. ONYX covers each stage at its confirmed content:
STAGE ONE: METR EVALUATORS
Amodei calls for embedded, third-party evaluators from METR — the Model Evaluation and Threat Research organization — to be granted permanent, employee-level access at every frontier AI company. This means evaluators would be physically present at the company, with access to the models, training runs, and safety documentation that company employees can access, not as auditors who visit periodically but as a continuous on-site safety function. Their mandate: verify safety practices and publicly report incidents. Amodei states Anthropic is ‘unilaterally committing’ to this regardless of what competitors do.
STAGE TWO: DEMOCRATIC THRESHOLD AGREEMENT
The second stage calls for democratic nations to agree on shared capability thresholds before advancing further. This means governments of democratic countries — not individual companies — would set the limits beyond which AI development should not proceed without additional safety verification. The specific mechanism: international agreement on what capability level requires what safety standard, modeled loosely on the arms control treaty framework.
STAGE THREE: THE CHINA DIMENSION
The third stage calls for extending the shared AI safety thresholds internationally, ‘including to labs operating in countries without democratic accountability structures.’ This is diplomatically careful language for a specific geopolitical reality: China is the world’s second-largest AI developer, operates without democratic accountability structures, and has not committed to any international AI safety framework. Amodei’s proposal requires that China and other authoritarian AI developers eventually be brought into the safety threshold framework. How that is achieved is the specific unresolved element of Stage Three.
WHY ALTMAN AND MUSK ENDORSED IT
Sam Altman (OpenAI CEO) and Elon Musk both publicly endorsed Amodei’s proposal within hours of its publication. The specific significance: these are Amodei’s competitors. OpenAI and xAI (Musk’s AI company) have competed directly with Anthropic. Their endorsement of a proposal that would impose third-party monitoring on their own operations reflects either genuine agreement on the risk, a calculated public relations alignment with safety messaging, or both. ONYX covers the endorsement as documented without adjudicating motivation.
Stage 1: METR evaluators inside every frontier lab, employee-level access, public reporting. Stage 2: democratic nations agree on capability thresholds. Stage 3: extend to countries without democratic accountability — which means China. That is the full architecture. Altman and Musk endorsed it the same day. The three most powerful AI company leaders are now publicly aligned on safety monitoring. What they do next is what matters.
WHAT HAPPENS NEXT
▸ METR evaluator implementation — whether Anthropic actually installs METR evaluators and whether competitors follow
▸ Democratic threshold agreement — whether any government initiates the Stage 2 international process
▸ China engagement — whether any diplomatic channel opens for China on AI safety thresholds
▸ Congressional response — whether the Altman-Musk-Amodei alignment produces legislative momentum
| CONFIDENCE: HIGH | Amodei METR evaluators employee-level access every frontier lab democratic nations capability thresholds countries without democratic accountability structures Altman Musk endorsement from confirmed CNN and essay reporting. |
SOURCES
▸ CNN / Amodei essay — three-step AI safety plan METR evaluators China September 14 2026
