↓ Skip to main content
sdd.sh
sdd.sh
  1. Articles/

Anthropic Says It Never Wanted an Open-Weights Ban — Here's What Amodei Actually Proposed Instead

·1071 words·6 mins·
Florent Clairambault
Author
Florent Clairambault
CTO & software engineer — writing daily about spec-driven development and agentic coding
Table of Contents

Anthropic Says It Never Wanted an Open-Weights Ban — Here’s What Amodei Actually Proposed Instead
On July 27, Dario Amodei published “Our position on open-weights models” on Anthropic’s newsroom, and it opens by directly rebutting a claim that’s been circulating for months: “Anthropic has never advocated for a ban on open-weights models.” Whether or not you believe that framing has been fair to Anthropic, the timing is not an accident. This is a company whose own blog has, in the last five weeks alone, covered Kimi K3’s 2.8-trillion-parameter open weights, GLM-5.2 beating GPT-5.5 at a sixth of the cost, Qwen3.8-Max’s unbenchmarked preview, and Anthropic’s own accusation that a 25,000-account distillation operation tied to Alibaba pulled 28.8 million API exchanges out of Claude between April and June. Anthropic is not a neutral observer in this debate — it’s the company whose models keep getting distilled, and whose export-control fights keep getting framed publicly as anti-open-source.

The Actual Position
#

Amodei’s post draws a distinction that’s easy to state and harder to hold consistently in a policy debate this politically charged: models “without dangerous capabilities” are, in his words, “a public good” — genuinely useful to businesses and researchers, and not something Anthropic wants restricted by default. The concern isn’t openness itself. It’s two specific downstream risks: authoritarian governments building superior AI capabilities, and powerful models enabling cyberattacks or biological threats. Both of those risks, Amodei argues, exist independent of whether a model’s weights are open or closed — which is exactly why a blanket open-weights ban would be the wrong tool even if you take the underlying concerns seriously.

Instead, the post lays out three specific proposals:

Chip export controls, enforced harder. Restrict the sale of powerful semiconductors and chipmaking equipment to China, and crack down on smuggling. This is not a new Anthropic position — export controls have been the backbone of the company’s policy asks since at least the Fable 5 restriction fight in June — but Amodei is explicit here that this is the primary lever for the national-security concern, not model licensing.

Targeted legal and commercial frameworks against distillation. This is the newest and most pointed proposal, and it reads directly against the backdrop of Anthropic’s own distillation accusation from a month ago. Amodei’s argument: industrial-scale distillation — extracting a frontier model’s capabilities through massive, systematic API querying rather than through legitimate research use — lets adversarial actors “partially evade chip bans” by borrowing capability instead of building the compute to train it themselves. His proposed fix isn’t restricting who can publish open weights; it’s going after the distillation operations directly, through legal and commercial mechanisms rather than blanket technical bans.

Mandatory safety testing, applied evenly. Any sufficiently capable model — open or closed, American or otherwise — should undergo pre-release testing for cyber, biological, and alignment risk before deployment. This is the part of the proposal that cuts against Anthropic’s own commercial interest most directly: it would apply to Claude the same way it applies to a Chinese open-weight release, and it explicitly rejects “closed models get a pass, open models get scrutiny” as a legitimate framing.

Why This Reads as Consistent Rather Than Defensive
#

It would be easy to write this off as damage control — a company under fire for its export-control advocacy trying to reframe the narrative after the fact. But laid against Anthropic’s actual behavior over the past two months, the position holds up better than a pure PR read would suggest. Anthropic didn’t ask Commerce to ban Kimi K3, GLM-5.2, or DeepSeek V4 from being used by American developers, even as this blog covered each of them approvingly for genuine open-weight capability gains. The company’s own Fable 5 export restriction — the 19-day suspension in June that this blog covered in real time — was explicitly framed by Anthropic as compute-and-capability containment aimed at a specific class of risk, not as a shot at open licensing generally.

The distillation crackdown proposal is the piece that most directly serves Anthropic’s commercial interest, and it’s worth being honest about that rather than pretending otherwise: a company that just accused a competitor’s affiliated actors of extracting 28.8 million API exchanges’ worth of training signal obviously benefits from legal tools that make that harder. But “obviously self-interested” and “wrong” aren’t the same thing, and the proposal itself — target the extraction method, not the publication of weights — is a materially different ask than “stop China from open-sourcing models,” which is the strawman the post is explicitly responding to.

The Broader Pattern This Fits
#

This blog has spent much of 2026 tracking a recurring tension: genuinely excellent open-weight models shipping out of Chinese labs (Kimi, GLM, Qwen, DeepSeek) at the same time American export-control policy has repeatedly constrained Anthropic’s own top-tier model (Fable 5’s June export ban, its subsequent billing whiplash, its eventual permanent-tier settlement in July). Reading those two threads side by side, the obvious question has been whether Anthropic’s regulatory advocacy is a coherent policy position or just protectionism dressed up as safety concern. Amodei’s post is the most direct attempt yet to answer that question in public, and the answer it gives — narrow the target to compute access and extraction methods, test capability regardless of license, leave open publication alone — is at least falsifiable in a way “we support safety” usually isn’t. The test now is whether Anthropic’s actual lobbying behavior over the next few months matches the position it just wrote down, particularly on the distillation-framework proposal, which doesn’t yet exist as actual legislation anywhere.

For developers building on open-weight models day to day, the practical upshot is simpler than the geopolitics: nothing here changes what you can legally use today. Kimi K3, GLM-5.2, and the rest remain available under whatever license terms their publishers actually attached (worth reading closely, per this blog’s own experience catching a mis-reported Kimi K3 license just two days ago). What Amodei’s post signals is where the regulatory pressure is more likely to land next — on the pipes moving capability out of frontier models via API, not on the open-weight ecosystem itself.

Sources:

Related