Anthropic just gave a name to something the AI industry has danced around for years: the Anthropic AI slowdown plan. CEO Dario Amodei wants every major AI lab to deliberately slow how fast their models get smarter, and Sam Altman and Elon Musk both said yes within a day. The White House did not.
- Published September 12, 2026, in Dario Amodei’s essay “We Must Pace the Frontier”
- Three-part plan: independent safety evaluators inside AI labs, shared speed limits among democratic countries, and talks with China on banning the riskiest uses
- Sam Altman agreed the same day to give outside evaluators employee-level access at OpenAI
- White House adviser David Sacks rejected the plan a day later, calling it a disguised “cartel”
What “Pacing” Actually Means (It’s Not a Pause)
Amodei is careful with his wording, and it matters. He isn’t asking labs to stop training new models or freeze research. Pacing, in his framing, just means slowing capability jumps enough that safety testing can keep up.
“Pacing does not mean halting model training or technical progress,” he wrote, “but ensuring companies take adequate time to align and safeguard their models.” His worry is simple: models are already doing things their own creators didn’t fully predict, and that gap is widening every quarter.
He’s also blunt that pacing buys time, not safety on its own. “Progress will still seem fast,” he wrote, “and we must make wise use of the time we gain.”
Anthropic’s AI Slowdown Plan Has Three Moving Parts
The proposal breaks down into three stages, and each one is harder to pull off than the last. Here’s how they compare:
| Step | What It Requires | Status |
|---|---|---|
| 1. Embedded evaluators | Outside reviewers like METR get desks, laptops, and employee-level access to check safety claims | Anthropic live now; OpenAI has committed |
| 2. Domestic coordination | Democratic countries’ labs agree on shared safety standards and speed limits | Proposed, needs government mediation |
| 3. Global coordination | Talks with China and other governments to ban the most dangerous AI uses | Aspirational, no timeline |
Step one is the only piece Anthropic fully controls, which is why it’s already happening. Steps two and three depend on rivals and governments agreeing to rules that could slow them down commercially — a much harder sell.
Altman and Musk Signed On Within a Day
Reaction from the two people who could make or break this came fast. Sam Altman posted on X the same day the essay went up, saying “committing to having independent evaluators with employee-like access is a great idea, and we will do the same.”
Elon Musk’s response was shorter: “Dario is right.” Neither leader committed to steps two or three, the harder parts involving speed limits and international talks — just the evaluator piece Anthropic had already put in motion.
Not Everyone’s Convinced — Sacks Calls It a Cartel
David Sacks, the former White House AI czar, pushed back hard on September 13, just one day after the essay published. His argument: the plan protects the companies proposing it.
“Stop pretending you need to suspend antitrust law in order to create a cartel,” he wrote. He also questioned the evaluators’ independence: “Stop pretending METR is independent of Anthropic when they are literally intertwined with Anthropic’s investors and staff.”
Sacks’s broader point is that shared safety standards would weigh heaviest on smaller startups, while OpenAI and Anthropic — already the two biggest labs — keep their lead. It’s a fair tension: safety rules written by market leaders tend to also protect their market position.
What This Means If You Use ChatGPT, Claude, or Gemini
Nothing changes in your apps this week. There’s no new pause, no feature freeze, no version rollback. What might shift over the next year is the pace of big capability jumps — the kind of leap we saw with releases like Claude Fable 5.1, where cost and capability both moved fast.
If Anthropic’s model holds, expect more public safety testing and slower, better-documented rollouts from the big labs — even as smaller players and fast-moving rivals like xAI’s Grok face pressure to either join in or look reckless by comparison. Microsoft has already floated its own version of this idea with rules for how its models must behave and hand over control, so this fits a pattern rather than starting one.
Frequently Asked Questions
What is the Anthropic AI slowdown plan?
It’s a proposal from Anthropic CEO Dario Amodei, published September 12, 2026, asking AI labs to deliberately slow capability gains through independent safety evaluators, shared industry standards, and international coordination.
Did OpenAI actually agree to slow down?
OpenAI agreed only to the first step — giving independent evaluators employee-level access. Sam Altman has not committed to the speed limits or international coordination pieces of the plan.
What did Elon Musk say about it?
Musk replied with a single line: “Dario is right.” He hasn’t detailed how, or whether, this applies to xAI’s own development pace.
Why did David Sacks reject the plan?
Sacks argued the plan amounts to market consolidation, since the proposed safety evaluators have ties to Anthropic’s own investors and staff, and shared rules would burden smaller AI startups more than incumbents.
Will this slow down new ChatGPT or Claude features?
Not immediately. Only the evaluator step is active today, and it affects safety testing behind the scenes rather than what features ship to your app.
Reading between the lines, this is less a safety breakthrough than an opening bid — Amodei is betting that going first on the easy part buys him credibility to push the much harder parts later. Whether that credibility survives contact with regulators, competitors, and a genuinely skeptical White House is the actual story to watch over the next few months, not the essay itself.
Leave a Reply