The Absence That Got More Attention Than Any Signature

In July 2026, an open letter started circulating in Washington defending open-weight AI models against any attempt at regulatory restriction. The list started at 25 companies and grew fast to more than 100 — Nvidia, Microsoft, Meta, Google, Palantir, and after a weekend of pressure, even OpenAI and Google signed on, both having originally sat out.

I noticed one absence before even finishing the list: Anthropic didn’t sign.

And the fact that Anthropic operates exclusively with closed models — the Claude family — made the accusation write itself. The pushback online was immediate: the company must be using “national security” as corporate cover to block open-source competition and protect its own business model.

I was suspicious too, at first. But once I dug into the full context, the story turned out to be a lot more interesting than “big company protects its market.”

The Scare That Started All of This

The crisis grew out of a real shock: Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter model that came close to matching the best American frontier systems, at a fraction of the cost.

That triggered panic in Washington. Reports circulated that the Trump administration was weighing a full restriction on Chinese open models, citing national security risk. The industry — led by Nvidia’s Jensen Huang — reacted fast, worried that this legitimate concern could turn into a blanket regulation capable of choking the entire open-source ecosystem, American companies included.

That’s the chain reaction the letter was born out of. And that’s the same climate that turned Anthropic’s absence into a headline.

What Anthropic Actually Said

Under pressure — including from a White House AI adviser — Dario Amodei published a piece titled “Our position on open-weights models” to lay out where the company actually stands.

His central point wasn’t “I’m against open-source.” It was more specific: Amodei has never advocated banning open-weight models as a category. What he disputes is the claim that open models necessarily help defense more than offense in cybersecurity — he argues that, in certain scenarios (like biological risk), the opposite could just as easily be true. He also raises a separate concern: Anthropic has been publicly accusing Chinese companies, including the maker of Qwen, of industrial-scale distillation of its models — a problem he thinks needs its own regulatory response, independent of the open-vs-closed debate.

Summed up the way I understood it: it’s not “open-source is dangerous.” It’s “every sufficiently capable model, open or closed, should go through mandatory safety testing” — which is a very different position from “ban open-weight.”

The Backfire (On the White House, Not on Anthropic)

Just days after that public exchange, the White House finalized its new AI safety guidelines framework — and the result caught everyone off guard, including Anthropic itself.

The framework, which grew out of a June executive order, is voluntary, but here’s how it works: only closed models that hit a “state-of-the-art” threshold in cybersecurity and hacking capability get invited into a government review window of up to 30 days before release. Open-weight models — American or Chinese, including the very Kimi K3 that kicked off this whole crisis — are exempt from that review, regardless of how capable they are.

Model categoryExamplesWhat the framework covers
Closed frontier modelsClaude (Anthropic), ChatGPT (OpenAI), Gemini (Google)Voluntary cybersecurity review window of up to 30 days before release, if they hit the “state-of-the-art” threshold
Open-weight modelsLLaMA (Meta), Kimi K3 (Moonshot)Exempt from review, regardless of capability

What I Actually Think

What strikes me most about this whole story isn’t who signed the letter and who didn’t. It’s the irony of who ended up footing the bill.

Anthropic — the company the community accused of trying to choke open-source to protect its own business — is precisely the one now carrying the heavier regulatory burden. Meanwhile, the open ecosystem, including Chinese models, remains free to be distributed with no prior review at all.

I think it’s too easy to reduce this to “Anthropic tried something and it backfired.” What I see is a genuine attempt to separate two debates the industry had mostly been treating as one: “open vs. closed” and “safe vs. dangerous” aren’t the same question. Anthropic pushed to keep that distinction, and the practical result was a framework that created exactly the asymmetry it said it feared — just pointed in the opposite direction everyone expected.

I’m Left With This Question

Regulation written under panic tends to hit the wrong target. In this case, the sector left standing was exactly the one that caused the initial panic — cheap Chinese open models — while whoever built a closed model with heavy investment in safety now carries the bureaucratic weight.

Do you think open models really should be exempt from government audits, or should regulation be identical for everyone, open or closed?

The company accused of trying to choke open-source is the one that ended up under government review. The open-weight models that scared Washington got exempted. Tech policy’s irony outran the very debate that started it.

Read Also