Anthropic's Amodei Says It's Time to Pace the Frontier, Not Race It

Pacing, not pausing

In a new essay called "We Must Pace the Frontier," Anthropic CEO Dario Amodei argues that frontier AI companies need to deliberately slow the rate at which they improve model capabilities. He's careful to distinguish this from a moratorium: "Pacing does not mean halting model training or technical progress, but ensuring companies take adequate time" to align and safeguard systems before pushing them further.

The essay lays out a three-step plan, moving from something Anthropic can do unilaterally today to something that would require the cooperation of governments that don't currently trust each other.

Step one: evaluators get a desk, not just an NDA

The concrete, near-term commitment is the most interesting part. Anthropic says it will give third-party evaluators — organizations like METR — ongoing, employee-like access: desk space, badges, company laptops, and tooling comparable to an internal risk team. Crucially, the arrangement is contractual: evaluators keep the right to publish findings on risk levels, incidents, and practices, and Anthropic can only redact narrowly for security, legal, or confidentiality reasons.

That's a meaningfully different model from the usual pattern of a lab commissioning a one-off external audit and choosing what to release. An embedded evaluator with standing access and publication rights is closer to how financial auditors or compliance monitors operate inside regulated firms.

The harder steps: coordination, then geopolitics

Step two asks frontier labs within democracies to agree on common safety standards and limits on how fast capabilities advance — potentially through checkpoints tied to what models can actually do, or caps on inputs like training compute. Amodei is explicit that this only works if the U.S. and allies don't fall behind China's pace, and pairs it with defensive measures: restricting AI chip exports, curbing model distillation, and hardening against model-weight theft.

Step three — talks with authoritarian governments, including China, on global AI pacing — is where Amodei's own confidence drops. He sketches four levels of possible agreement, from narrow bans (no AI-assisted bioweapons work) up to capping the speed of recursive self-improvement, and says full global pacing is "unlikely to actually happen any time soon," though narrower deals show "some edge of being possible."

We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast.

— Dario Amodei, "We Must Pace the Frontier"

Why builders should care

If you're shipping products with an LLM in the decision loop — a trading agent sizing positions, a smart-contract assistant generating Solidity, a support bot with tool access — the direction here matters more than the geopolitics. The industry's most cited safety voice is normalizing the idea that:

• External review with real system access, not marketing-copy "red teaming," becomes a credibility signal for AI vendors and the products built on top of them.
• Capability growth and safety review are being framed as sequential, not parallel — expect model providers to lean harder on staged rollouts and capability evals before wider release.
• Teams building on frontier models should expect more provenance and incident-reporting expectations to trickle down from labs to API consumers over time.

None of this changes what you ship this week. But it's a reasonable bet that "can you show your safety evaluation" becomes a question enterprise buyers ask AI-driven fintech and trading tools, the same way they ask about SOC 2 today.

FAQ