What Would AI Labs Do With a Slower Schedule?

1 The proposal on the table

Anthropic chief executive Dario Amodei published an essay in September 2026, 'We Must Pace the Frontier,' proposing that frontier labs deliberately slow their release cadence. It is a proposal. No lab has adopted it and no regulator requires it. Amodei, who co-founded Anthropic in 2021, frames the slowdown as a gift of time — which raises the question of what the time is for.

2 The safety-testing use of time

According to Amodei's proposal, additional months between frontier releases would be spent on deeper pre-deployment testing: longer adversarial evaluation of agentic behavior, wider red-teaming, and fuller measurement of misuse potential before the public ever touches a system. The implicit claim is that today's evaluation windows are too short for each frontier capability jump.

3 The research use of time

The proposal also allocates time to interpretability and control research — understanding what a model is doing internally and building reliable ways to correct it. Amodei's essay argues that alignment techniques lag capabilities because capability research is better incentivized. A slower schedule, in his framing, would let the lag shrink without a capability freeze. Whether the lag would actually shrink is an assumption of the proposal, not a demonstrated result.

4 How progress measurement would change

A slower frontier changes what progress means. Under current practice, progress is benchmark gains shipped on a steady beat. Under the proposal, progress would be measured by the maturity of safety systems — evaluation coverage, interpretability depth, incident-response readiness — rather than by release velocity alone. The essay proposes, in effect, a different scoreboard, and the scoreboard is the part outsiders can see.

5 What outsiders could verify

Release dates are publicly observable; internal safety investment is not. Verification of the proposal's commitments would have to lean on the artifacts that leak into public view: published evaluation results, third-party audit summaries, disclosed incident reports, and the cadence of releases themselves. Commitments about research effort inside the lab would remain claims unless paired with outside inspection.

6 The competitive constraint

The proposal's stated weakness is that pacing binds only those who adopt it. A lab that slows while rivals accelerate would spend its time well and lose the frontier anyway — a tension the essay acknowledges by calling for collective action. No proposed mechanism yet makes the slowdown reciprocal.

7 The central question, answered

What would labs do with a slower schedule, and how would anyone know? Per the proposal, they would test longer, research alignment and interpretability, and report against a safety-maturity scoreboard. Outsiders could verify the cadence and the published artifacts, but the core commitment — how the extra time is actually spent — would rest on lab claims until inspection regimes exist.