Anthropic: Open Weights Should Not Be Banned

What Exactly Anthropic Stated

The main point here is simple: Anthropic is not calling for a ban on open-weight models as a class. In its official material, the Position on Open Weights Models, the company writes the opposite: models without dangerous capabilities are a public good, and regulation should target specific risks.

What matters most to me is not the phrasing about open weights itself, but the framework. Anthropic draws the line not between open and closed, but between relatively safe models and frontier-capable systems that can amplify misuse.

In the same document, Dario Amodei names three measures the company considers reasonable: stronger chip export controls, action against industrial-scale distillation, and mandatory safety testing for sufficiently capable models, whether open or closed.

And here the position becomes much clearer. If before the debate often devolved into slogans like "open good" or "open dangerous," Anthropic formulates a more engineering criterion: we look at capability, access, and misuse, not the ideology of licensing.

It’s also telling that this document contains no new benchmark or flashy number, which usually disguises a political message. This is not a model release; it’s an attempt to redraw the debate map so that regulators have a set of thresholds and triggers, not a club.

Why This Changes the Conversation

In short: this is a blow to the idea of a blanket ban and simultaneously supports stricter policies where a model is truly dangerous. This stance is convenient for developers of open-weight models and for regulators who need a working rule, not a slogan.

First practical implication: the debate can no longer be reduced to "are you for or against openness?" If we accept Anthropic’s framework, we have to discuss which capabilities are considered dangerous, how to measure sufficient capability, and who conducts safety testing before release.

Second practical implication: the focus is no longer only on publishing weights but also on the compute chain. Chip export controls and the fight against industrial-scale distillation mean that policy will revolve around compute and capability copying, not just model licensing.

My conclusion is boring but important: this is neither a capitulation to the open-weight camp nor an attack on it. It’s an attempt to establish capability-based regulation as the language of the whole discussion. And going forward, the most uncomfortable question will not be "to open or not to open," but who decides, and by which tests, that a model has crossed a dangerous threshold.

Previously, we discussed how Anthropic restored transparency after the scandal, canceling hidden quality downgrades for queries. This step is directly tied to their current stance on open-weight models.