Breaking

Anthropic CEO Dario Amodei Calls for Slowing Frontier AI as the Safety Lab He Built Faces the Race It Was Designed to Resist

September 13, 2026 at 05:57 EDT

Foundation Models
Policy & Regulation
Funding & M&A

Dario Amodei left OpenAI in 2021 to found a laboratory insulated from a commercial sprint he considered unsafe. On Saturday he argued that the sprint has arrived anyway, and that frontier companies — including Anthropic — must deliberately slow how fast they improve model capabilities so safety work can keep up.

September 2026 · Anthropic

The safety lab meets the race

Dario Amodei built Anthropic to resist a commercial sprint he considered unsafe. He now says that sprint has arrived, and that frontier labs must slow capability gains so safeguards can catch up.

6–12 mo

Soonest window he cites for a misaligned swarm seizing the internet

1–2 yr

Extra time he says could cut the chance of a serious failure

~2%

Amodei’s equity after years of fundraising; founders seek super-votes

3 steps

Evaluators first, then democratic coordination, then global limits

Warning window vs. time sought

The clock he cites is shorter than the pause he wants

6 mo

Soonest swarm risk he named

12 mo

Shortest extra year he said could help

Takeaway: the minimum warning window is half the minimum extra year he argues safety work needs.

How the plan is supposed to work

Three steps, only the first is unilateral

Step 1

Embedded evaluators

Permanent, employee-like access. Anthropic says it will do this now.

→

Step 2

Democratic coordination

Shared safety standards among labs in democracies, likely with a government waiver.

→

Step 3

Global coordination

Attempt verifiable limits with authoritarian governments, without losing a U.S. lead.

Support

Rivals say they will match the first step

Sam Altman called employee-like evaluators a great idea and said OpenAI would do the same. Elon Musk backed the core argument. Hugging Face asked to join the evaluator effort.

Pushback

Others call the warnings hype or a power play

Some executives treated extinction talk as marketing around a listing. Critics said the plan would concentrate power. The White House has rejected extinction fears and focused on beating China.

The constraint he will not drop

Pacing is not a halt

Amodei says companies should still train models, but take enough time to align them — and keep the U.S. lead over China. That speed limit is the same tension the lab was founded to manage.

No full stop on training Preserve commercial advantage Do not cede lead to China

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.