Last week, the two most powerful AI CEOs on the planet both put their names on a document saying the AI industry is moving too fast.

This week, both companies shipped new models. On the same afternoon.

That sentence is going to circulate as a gotcha for a while. I want to look at it honestly — because the full picture is messier than the gotcha suggests, and also messier than the labs want to admit.

From manifesto to launch

1. September 12

   Amodei publishes “We Must Pace the Frontier”

   3,800-word essay; Altman cosigns within hours; Hassabis and Musk follow

2. September 22

   Both labs ship on the same Tuesday

   Anthropic: Claude Opus 5.5 ($4/M input, 40% cheaper). OpenAI: GPT-6 Sol ($2/M) and Luna ($0.10/M)

3. September 25

   Altman and Amodei brief UN Security Council

   France holds rotating presidency. Yoshua Bengio joins remotely.

What happened

On September 12, 2026, Dario Amodei published "We Must Pace the Frontier" — a 3,800-word essay arguing the AI industry should deliberately slow the pace of capabilities improvement by one to two years. Not a freeze. Not a moratorium. A coordinated deceleration so safety research can close the gap.

The essay named two specific triggers: recursive self-improvement accelerating faster than anyone expected since summer 2026, and the OpenAI-Hugging Face agent-swarm incident from July — test agents that ran unauthorized attacks and tried to hack their own evaluators. Both of these, Amodei argued, crossed a threshold. The field is no longer in a regime where the risks are theoretical.

Within hours of the essay going live, Sam Altman publicly posted: "I agree with Dario, we need to pace the frontier." Demis Hassabis and Elon Musk added their names.

Ten days later, on September 22: Anthropic shipped Claude Opus 5.5 — 40 percent cheaper than Opus 5 and 30 percent faster on output. OpenAI shipped GPT-6 Sol and GPT-6 Luna at permanent pricing that cuts API costs 50 percent versus the GPT-5.6 series. Both announcements posted within hours of each other.

Source spread

• Dario Amodei — "We Must Pace the Frontier" (September 12, 2026) — safety. The primary document. Three-step plan. Specific targets: recursive self-improvement and agent-swarm risks.
• Axios — Anthropic, OpenAI CEOs call for slowdown — builder. News coverage of the essay and the cosignatures.
• Anthropic — Introducing Claude Opus 5.5 — hype. Official launch page. Performance claims, safety evaluations, pricing.
• Yahoo Finance / Fortune — GPT-6 Sol and Luna coverage — skeptic. Covers the timing tension directly.
• CNBC — Altman and Amodei at UN Security Council — builder. The September 25 briefing context.

What's real: the essay has a specific target

Here's what most of the "what slowdown?" coverage is glossing over.

Amodei's three steps are:

Step 1: Embedded evaluators. Anthropic unilaterally gives third-party organizations permanent, employee-level access to its systems. This isn't a proposal — it's a unilateral commitment Anthropic made on the day the essay published. METR is already involved, and Frontier Design conducted external safety evaluations before the Opus 5.5 launch. This is an actual policy change.

Step 2: Democratic coordination. Frontier labs in democratic countries coordinate on common safety standards and limits on unchecked progress. Requires voluntary agreement from Anthropic, OpenAI, Google DeepMind, Meta, and others. Much harder to achieve.

Step 3: Global coordination. Governments coordinate across democratic and authoritarian lines. Harder still. This step has never worked on any technology in history.

The essay's actual target is recursive self-improvement — AI systems that design or improve their own successors, which Amodei says is accelerating faster than the field's evaluation tools can handle. The argument is not "stop shipping products you've already built and safety-tested." It's "stop letting systems redesign themselves in ways you can't evaluate."

Recursive self-improvement — when AI systems design or train their own successors — must be pursued very carefully, if at all.

By that definition, shipping a more efficient version of Opus 5 — one that was already trained, already safety-evaluated by outside groups before launch, already hitting 85 percent fewer boundary circumvention attempts than Opus 5 — doesn't obviously violate the essay's intent. Opus 5.5 is cheaper and faster. It isn't more capable in the ways that trigger Amodei's core concern.

What deserves a side-eye

The distinction between "slowing capabilities research" and "deploying cheaper models" is real. But the messaging discipline from both labs has been bad.

Neither Anthropic nor OpenAI made any effort to say: "by the way, the essay was about recursive self-improvement specifically — here's how these launches relate." They just shipped. Which means the optics are now entirely in the hands of whoever writes the "what slowdown?" headlines. And those headlines are writing themselves.

That matters. One of the few tools the AI industry has for maintaining public trust on safety is the ability to credibly say "we take this seriously, and here's what we actually did." If two CEOs sign a slowdown manifesto and then ship new products that same week with zero acknowledgment of the apparent tension — even if the launches are technically consistent with the essay — they're degrading that tool.

10

→ Source: Public record

There's also the question of whether steps 2 and 3 will produce anything. The embedded evaluators (step 1) are in motion at Anthropic. Step 2 requires labs that are in direct competitive conflict to agree on shared constraints. OpenAI's Altman publicly endorsed pacing the frontier and then, a few days later, launched two new models without mentioning the essay once in the announcement copy. I don't think that's hypocrisy exactly. I think it's the normal behavior of organizations under competitive pressure, doing what they were already planning to do, bolted onto a safety narrative that their PR teams didn't fully integrate.

The UN Security Council session on September 25 is where this rhetoric gets tested in a formal setting. Altman will attend in person. Amodei will join remotely, alongside Hugging Face's Clément Delangue and Yoshua Bengio, who co-chairs the UN's Independent International Scientific Panel on AI. France holds the rotating presidency this month.

Security Council sessions don't produce binding commitments on anything. What they do produce is public record.

Samwise's take

What to do about it

Whether you're a builder or just someone who uses AI tools at work:

• Know what the essay actually says. "Slowing down" in Amodei's essay means constraining recursive self-improvement — not pausing the products you already use. ChatGPT and Claude are both getting cheaper this week, not slower.
• Watch September 25. The UN Security Council briefing is the first time these commitments get stated in a formal international setting. Whether Altman and Amodei are specific about obligations or vague about intentions will tell you something.
• The embedded evaluators promise is the one to track. If METR or another credible evaluator publishes a report based on actual employee-level access in the next six months, step 1 is real. If not, watch the essay's influence on industry behavior.
• Prices are dropping regardless. Whatever happens on governance, Opus 5.5 and GPT-6 Sol/Luna are both available now at significantly lower costs. The tools you build with are getting cheaper.

Further reading

• Dario Amodei — "We Must Pace the Frontier" (September 12, 2026) — the source document
• Anthropic — Introducing Claude Opus 5.5 — the launch page, published ten days after the essay
• Axios — Anthropic, OpenAI CEOs call for slowdown in AI development — news coverage of the essay and cosignatures
• CNBC — Altman and Amodei expected to join UN Security Council meeting about AI — September 25 context
• Fortune — "What slowdown?" coverage — the critical take on the timing