The Safety Door and the Paywall Are the Same Door

Anthropic built its most powerful model and then refused to let the public use it. Claude Mythos, the company said, could find software vulnerabilities on its own, a capability too dangerous to hand to anyone. So the model went to Project Glasswing instead, a consortium of more than 40 organizations that includes Microsoft, Google, Amazon, Apple, and JPMorgan ^[1]. The public gets nothing. The biggest companies in the world get the strongest model Anthropic has ever made. Sam Altman and David Sacks did not buy the safety argument.

AI will probably most likely lead to the end of the world, but in the meantime, there'll be great companies. — Sam Altman

The rivals are naming something on the record: from outside the lab, a safety restriction and a commercial gate look like the same thing. And the architecture keeps repeating the same shape. This month Anthropic published GRAM, a method that isolates dangerous knowledge inside a model into discrete modules that can be switched on or off, so developers can decide who gets which capabilities ^[2].

Frontier AI models have knowledge that could be misused for nefarious purposes. — Anthropic

That is a safety feature. It is also a dial that can be turned to create paid tiers of capability, and nothing in the design says which way it is being turned. The same shape shows up in the fine print. Anthropic revoked OpenAI's API access after discovering OpenAI staff had used Claude to refine GPT-5, permitting only benchmarking and safety testing while banning model development ^[3].

While we respect Anthropic’s decision to cut off our API access, it’s disappointing considering our API remains available to them. — OpenAI

A rule that keeps a dangerous model out of a rival's hands is also a rule that keeps a rival from building on your work. The honest counter is that the safety concern is not invented. Mythos really can find zero-day vulnerabilities. OpenAI has released open-weight models, and California judges ruled that training on books is fair use ^[4][5]. The safety motive is real. The commercial motive is real. What no one outside the lab can do is say which one drives which gate. The same door serves both purposes, and the labs' own competitors have stopped trying to tell them apart.