Why Anthropic wants the whole AI industry to slow down (except itself)

Picture this: the CEO of one of the world's most powerful AI labs publishes a essay begging the entire industry to slow down. Within a day, it rackes up tens of millions of views. Rival CEOs publicly agree. And then, less than a week later, that same CEO's company ships a massive new product redesign, rebuilds its flagship coding feature, and rolls out a smarter reasoning mode — all while the "let's slow down" essay is still trending.

That's not irony. That's just how the AI industry works right now. And if you want to understand where things are headed, this week's news is a perfect case study.

What did Dario Amodei actually say in his "pace the frontier" essay?

On September 12, 2026, Anthropic CEO Dario Amodei published an essay calling on the few labs at the frontier of AI, the ones building the most advanced models, to slow down together. You can read the full thing on Dario Amodei's personal site, where he lays out a three-part plan.

The core idea isn't "stop building AI." It's closer to "let's not sprint blindly." As one breakdown of the essay put it, "We Must Pace the Frontier" is Dario Amodei's essay that asks frontier labs to slow capability growth, not to stop building AI. Amodei's own framing is even more direct: "We must slow the pace at which we improve the capabilities of AI models." But he's careful to clarify that pacing does not mean pausing — progress will still seem fast.

The proposal breaks down into three steps, and Anthropic didn't just talk about the first one — it committed to it. According to Amodei's own post on X, Anthropic is unilaterally committing to providing third-party evaluators with permanent, employee-level access to its systems, so that they can verify adherence to safety measures, report on incidents, and assess models' alignment during training.

The second proposal is what Amodei calls Democratic Coordination — frontier AI companies within democratic countries coordinating to establish common safety standards as well as limits on the rate of unchecked AI progress. He's upfront that this part is legally messy. Some forms of coordination that would be impactful for pacing are legally challenging, and will require government support.

Who actually agreed to slow down, and who didn't?

Here's where it gets interesting. The essay didn't land in a vacuum — it landed with immediate reactions from basically every major player in AI.

On the "yes, let's do this" side: OpenAI followed Anthropic's lead, and both Elon Musk and Demis Hassabis endorsed the overall proposal. A separate roundup of reactions confirms the list ran even wider — publicly supporting step one, embedded evaluators, were Amodei, Altman, Musk, Hassabis, and Nadella. Microsoft's Satya Nadella reportedly added his own condition, arguing that closed and open models should both keep a seat at the table.

But not everyone bought it. Critics pointed out the awkward incentive structure baked into the proposal — namely, that a voluntary slowdown mostly benefits the labs that are already ahead. As one analysis noted, the costs of pacing are easiest to carry for the labs with the deepest pockets, so a voluntary slowdown effectively taxes the challengers more than the leaders. Meanwhile, open-source and Chinese labs are closing the gap on the leaders — Alibaba's Qwen models and DeepSeek keep putting out new versions fast, none of which are bound by any pacing agreement.

The pushback got louder outside the industry too. Reports circulating a few days after the essay noted that investor Michael Burry called the pacing warnings "self-serving hype tied to IPOs," while other public figures weighed in from opposite political directions. Not exactly the unified "let's all be careful together" moment Amodei might have hoped for.

So what did Anthropic actually ship right after calling for a slowdown?

This is the part that makes the whole story worth paying attention to. Just days after "pace the frontier" went viral, Anthropic rolled out one of the biggest structural changes to Claude in months — a full redesign of Claude Projects.

You can read the official announcement directly on Anthropic's blog, and the technical documentation lives at Claude Code Docs. The headline says it all: the company's blog post is titled "Projects redesigned: from folder to conversation" and describes "a new experience for Claude projects, now available in beta in Claude Code."

The change is bigger than a UI refresh. Previously, Projects functioned as a static context folder that bundled together reference files — prompts, design documents, API specs — while breaking work into tasks and handing off context between sessions all fell to the developer. The new version flips that. According to Anthropic's own post, projects now have threads that do the work and a coordinator that directs them — you select a goal as well as the repo or context, and Claude starts by suggesting work it can pick up right away.

Here's a concrete example straight from Anthropic's launch post: you can configure a project and set a goal to reduce your app's checkout p75 latency, then ask Claude to profile each endpoint, test optimizations, and open PRs in parallel threads. That's not a chatbot answering questions — that's a coordinator managing multiple parallel coding sessions on its own.

If you want to try it yourself, here's how the rollout actually works right now:

1. Go to claude.ai/code or open the Claude desktop app and check the Code tab.
2. Availability depends on your plan and usage. Per the docs, projects are in public beta on Pro and Max plans and rolling out gradually, starting with accounts that have used cloud sessions and don't have existing projects in claude.ai chat or Cowork — they aren't available on Team or Enterprise plans yet.
3. If you don't see it, Projects won't appear in the sidebar, which means the rollout hasn't reached your account yet, and you can join the waitlist directly through Anthropic's site.
4. If you already have existing projects, don't worry — existing projects on Pro and Max plans keep working as they do today, and Anthropic will upgrade them as the rollout expands to chat and Cowork.

What does "extended thinking" have to do with any of this?

The other piece of this week's puzzle is Claude's reasoning upgrades, which tie directly into the "coordinator" model above. Anthropic's official documentation on extended thinking explains the underlying mechanism well: it's not a different model — it's the same model given more time and effort to come to an answer, which gives it a real boost in intelligence.

What's changed recently is how that reasoning gets exposed to users and developers. Newer models have moved toward a smarter default: thinking is now summarized by default, so you get the key reasoning steps instead of the raw token stream, and there's a new adaptive thinking mode that replaces the fixed budget-tokens approach. That adaptive approach means Claude essentially decides for itself how hard to think about a given problem, rather than a developer manually setting a token budget every time.

For anyone building real products on top of Claude, Anthropic's own guidance is blunt about when this matters: use extended thinking for particularly complex tasks that benefit from step-by-step reasoning like math, coding, and analysis. That's exactly the kind of workload the new Projects feature is designed to hand off to Claude in parallel threads — deep reasoning, running unattended, across multiple sessions at once.

Why does Apple's Siri saga feel so out of step with all this?

Zoom out to Apple's side of the AI world and the contrast is almost comedic. While Anthropic and OpenAI are racing to ship coordinator agents and adaptive reasoning, Apple is still working through a multi-year plan just to get a competent AI assistant into Siri.

The current plan has Apple leaning on Google's technology, with Google confirming that Gemini will power Apple Foundation Models going forward, after Apple determined Google's technology provides the most capable foundation. A more conversational, chatbot-like Siri isn't expected until the iOS 27 rollout, and reporting has pegged it as launching a more thorough redesign of Siri as a full chatbot, unveiled at WWDC on June 8, 2026, with a general release in September 2026.

So while frontier labs are publicly debating whether they're moving too fast, Apple is still trying to catch up to where those labs were over a year ago. It's a useful reminder that "the AI industry" isn't one speed — it's several very different races happening at once.

What should you actually take away from all this?

The pacing debate is genuinely important — coordination between labs, third-party evaluators, and safety standards matter a lot if AI capabilities keep compounding. But don't mistake a safety essay for an actual slowdown in what you can use today. In the same week Anthropic asked the industry to pace itself, it shipped a coordinator-style Projects redesign and rolled out smarter adaptive reasoning. That's not hypocrisy so much as it is the current reality of AI development: the conversation about caution and the pace of shipping are running on two completely separate clocks.

If you're building anything on top of Claude right now, the practical move is to actually test the new Projects beta and see whether handing off parallel, coordinated work changes how you use the tool day to day. The essays about slowing down are worth reading. But the tools aren't waiting around for the debate to finish.