The safety brand and the lobbying position are pulling in opposite directions

Anthropic sells itself as the safety-first frontier lab. According to the New York Times, it is also privately lobbying the White House to restrict the models it competes against. Those two positions can both be sincere. They cannot both be costless, and this week the cost showed up as a public split inside the US tech industry.

The reporting. The New York Times, in a piece by Mike Isaac, Kate Conger, Ana Swanson and Meaghan Tobin published 25 July 2026, reports that OpenAI and Anthropic have lobbied Washington regulators to restrict open-source AI models, citing five people close to the discussions.¹ The Times reports that Treasury Secretary Scott Bessent and Michael Kratsios, President Trump's science and technology adviser, have weighed in on how American AI models are, in the paper's paraphrase, valuable intellectual property.¹ I have not read the NYT piece directly, it is paywalled and the research pack quotes it via secondary reporting, so I am treating it as reported, not proven, throughout.

The counter-move. On 24 July, 25 companies including Nvidia, Microsoft, Meta, IBM, Hugging Face, Palantir, Mozilla, the Linux Foundation, Andreessen Horowitz and Y Combinator signed a letter titled "Open Weights and American AI Leadership," urging Washington not to impose premature restrictions on open-weight AI models, according to AI Weekly.² OpenAI, Anthropic, Alphabet and Amazon were absent at publication. AI Weekly and PitchBook News report that OpenAI added its signature on Friday afternoon, after the initial omission was widely noticed on X.²³ Anthropic did not sign and, per PitchBook, did not respond to requests for comment.³

Why now. Open-weight releases from Chinese labs have moved from curiosity to competitive threat over the past year. AI Weekly, citing VentureBeat, reports that Moonshot AI's Kimi K3, released mid-July 2026 and backed by Alibaba, is a 2.8-trillion-parameter open-weight model that Moonshot benchmarks as comparable to Anthropic's and OpenAI's proprietary systems on several tasks, priced at $3 per million input tokens and $15 per million output tokens.² Those API prices sit below what closed frontier labs charge, and, more importantly, an open-weight model can be pulled onto private infrastructure and run at whatever the operator's inference cost is. The pricing sheet is the least of it.

The distillation frame. The policy argument for restricting open weights has ridden on the claim that Chinese labs are distilling Western frontier models, training smaller systems on the outputs of larger ones, to leapfrog the compute frontier. PitchBook News reports that researchers dispute this in the Kimi K3 case: Fable, the Anthropic model Moonshot is accused of having distilled, was not publicly available long enough for Kimi to have trained on it, according to those researchers.³ If the specific factual predicate is wrong, the national-security rationale attached to it is doing less work than it appears to.

What Amodei has said. AI Weekly reports that Dario Amodei has called open source "a red herring" and, in earlier congressional testimony, argued that scaling open-source models is going down a dangerous path.² PitchBook attributes to him the further claim that open source does not work the same way in AI as in other software domains.³ I have not seen a primary transcript for either statement; both are quoted in secondary reporting and I am flagging them as such. The direction of his public position, though, is not in dispute — Anthropic's chief executive has argued the open-source-risk case in public, repeatedly, for over a year.

Where the split gets structurally interesting. Microsoft is on the letter. OpenAI, in which Microsoft is the largest outside investor, was not — until it was. That puts Microsoft on record, publicly, against the lobbying position its portfolio company is reportedly running privately in Washington. Nvidia is also on the letter; PitchBook reports Nvidia has pledged $26 billion toward its own open-weight models and has invested in Hugging Face's $235 million round in 2023 and Reflection AI's $2 billion round.³ Chip demand for inference workloads runs through open weights as much as closed ones. The commercial interest is real and it is not hidden.

The regulatory-capture reading. David Sacks, per AI Weekly, has accused the closed labs of pursuing regulatory capture — using national-security framing to raise barriers against open-source rivals.² AI Weekly also paraphrases Suhail Doshi arguing that American labs trained on humanity's data without paying for it before lobbying to ban open-weight models in the name of distillation.² Those are pointed accusations from interested parties, and the pattern they describe, incumbent invokes safety, regulator raises barrier, competitor is priced out, is the textbook shape of the phenomenon they are naming. That does not automatically mean it is what is happening here. It does mean the burden of proof sits on the lobbying side, not the letter side.

The honest counter-case. Amodei was making the open-source-scaling argument before Kimi K3 existed in its current form. A prior that happens to align with commercial interest is still a prior. It is possible to believe, in good faith, that releasing frontier-capable weights into an ecosystem you cannot police is a bad idea, and that this belief becomes more urgent, not less, when the actor releasing them is state-adjacent. The problem is not that the argument is unavailable to Anthropic. The problem is that the company is making it in a Washington meeting room while declining to make it, on the record, in a letter signed by every other serious player in the US stack.

What has not happened. No draft executive order or piece of legislation restricting open-weight models has been publicly confirmed. The lobbying, if it is happening at the scale the New York Times describes, has not yet produced a policy instrument. Editor's note on a claim I have cut from an earlier version of this piece: an aggregator summary attributes to PBS a line about the White House seeing no need to restrict open-source AI "for now," but the underlying PBS reporting appears to be from 2024, under a different administration, and separate reporting suggests the current White House has not ruled action out. I do not know the current administration's position on this, and neither, on the evidence, does the coalition that just wrote the letter.

The signature list changed once already this week under public pressure. The next thing to watch is whether Anthropic's does, and, more consequentially, whether anything moves in Washington beyond meeting rooms.

Footnotes