Anthropic Finally Has an Opinion on Open Weights. It Has Never Shipped Any.

Anthropic published its official position on open-weight models this week. Within a day, the Hacker News thread hit 397 points and 529 comments. When comments outrun points by that margin, it means people weren’t reading — they were waiting to argue.

One caveat up front. What follows draws on Anthropic’s own document and a single Hacker News discussion. There’s no meaningful Reddit thread yet, and the broader industry hasn’t had time to form a considered take. So read this as a snapshot of the first developer reaction, not as a verdict from the field.

The Lab That Never Opened Anything

Meta shipped Llama. Mistral shipped weights. Alibaba’s Qwen shipped weights. DeepSeek shipped weights. Even OpenAI, after years of the joke writing itself, eventually released a small open-weight model.

Anthropic has released exactly zero. Not once since founding. No version of Claude’s weights has ever left the building. By any reasonable measure, it is the most closed frontier lab in the business.

So when that company publishes a document titled with its position on open-weight models, the reflexive suspicion is understandable. When a company that has never done a thing writes an essay about that thing, the essay is rarely a compliment.

The Core Claim: We Never Asked for a Ban

The line doing the most work in the document is this one: Anthropic has never advocated banning open-weight models.

That’s a rebuttal, not a policy. Anthropic has spent two years fielding the accusation that its regulatory lobbying squeezes the open-source ecosystem — a grievance that hardened during California’s SB 1047 fight. Anthropic backed safety rules premised on frontier-model risk. The open-source camp read those same rules as a noose sized for their necks.

Which makes this less a policy announcement than reputation management. Documents that open by denying something you said tend to be.

Hacker News Wasn’t Buying It

The thread was cold. The most-echoed sentiment, roughly: as expected, they’ll use the government to kill competitors.

Blunt, but the mechanism is not mysterious. When a frontier lab endorses safety regulation, the only entities that can absorb the compliance cost are frontier labs. Standing up a compliance team, running eval pipelines, staffing people to sit in rooms with regulators — all of that costs real money. Anthropic has it. Three grad students with a GitHub repo do not.

Sincere intent and captured outcome are not mutually exclusive. Economists call this regulatory capture, and it needs no conspiracy to operate. It’s just what happens when the cost of a rule scales differently than the ability to pay it.

The Real Sore Spot: The Distillation Clause

The objection that surfaced again and again in the comments was the ban on distillation. The word people kept reaching for was hypocritical.

Quick refresher on the technique. You take a large model’s outputs and use them to train a smaller one — the student watches the teacher work and learns to imitate it. DeepSeek’s low-cost, high-performance models are the best-known case where this was widely suspected.

Anthropic’s terms of service prohibit using Claude’s outputs to train a competing model. The friction is obvious: Anthropic, like every other frontier lab, trained on internet text scraped without asking anyone. Learning from other people’s copyrighted work is fair use. Learning from my model’s output is a terms-of-service violation. That’s the asymmetry the thread kept circling.

Legally, these are separate questions — copyright law and contract law genuinely are different instruments, and the distinction is not a dodge. But when the question is moral consistency rather than legal exposure, the answer gets thin fast.

The Case Anthropic Actually Has

Before piling on, the other side deserves a fair hearing.

The underlying worry about open weights is not invented. Once weights are public, they cannot be recalled. Stripping safety training via fine-tuning takes hours, not months, and uncensored variants of Llama-family models are sitting in public repos right now, easily found. An API model can be cut off mid-abuse. A downloaded checkpoint cannot.

Whether that concern is heartfelt or merely convenient is not something anyone outside the company can adjudicate. Worth holding onto, though: both can be true at once. A position being sincere and a position being profitable are not contradictory states.

What to Watch Instead

The document contains no actual change in policy. Anthropic will keep its weights closed. It will keep backing safety regulation. The only thing that shifted is the framing.

Which leaves the question that actually matters: when the company talking loudest about AI safety is also the company holding everything tightest, how much of that talk do you buy? The position paper won’t settle it. The next regulatory fight will — watch which specific clauses Anthropic supports, and which ones it quietly opposes. Positions are cheap. Line-item votes are not.