Ben Thompson: Anthropic's safety story is also its power story

The thesis: safety as strategy, not just conscience

In Anthropic's Safety Superpower, Ben Thompson argues that the company's safety stance gives it a rare alignment between "talent and mission and business" — and that this is precisely why it's so powerful and so worth scrutinizing. The events he strings together: a Mythos preview in April 2026 announced as too dangerous to release for its cyber capabilities; Fable shipped two months later with guardrails that were quickly jailbroken; and then U.S. ITAR export controls that forced access restrictions on both. His uncomfortable conclusion, quoted directly:

"Anthropic believes that they are the ones who should have final say over how Anthropic is used; given that they think only they should be developing leading edge AI, they by extension think that only they should have final say over AI generally."

Thompson's worry isn't that the safety belief is fake. It's that a sincere belief, fused to commercial and political incentives, "excuses decision after decision."

The thread takes the bait — both ways

The line that detonated was the one about scope of ambition.

"they by extension think that only they should have final say over AI generally… Anthropic's leadership effectively wants to have power over everything and everyone." — an engineer flagging the most troubling passage

The strongest counter went straight at the internal logic:

"You can't be on your way to 'power over everything' and get distilled into free Chinese models within months. Pick one. The bottleneck is compute and data, not the model." — a skeptic of the gatekeeper framing

His evidence is the ITAR episode itself: with no nationality controls actually in place, the only lever Anthropic had was killing access entirely — "not exactly what an all-powerful gatekeeper does."

A defender, and a third reading

Not everyone bought Thompson's frame. One reader drew a narrower line: Anthropic "guards their tools from mis-use… they are not trying to control anything or anyone," they just decide what counts as misuse. Another rejected the whole 'they-really-believe-it' defense as a blank check for any decision. And a director of engineering offered a data point cutting the other way: Anthropic's models top a benchmark measuring whether a model pushes back on nonsensical requests instead of hallucinating agreement — evidence the safety posture shows up in behavior, not just press releases.

When the moat and the morals are the same wall

The argument that lands is that you can't cleanly separate Anthropic's convictions from its competitive interest, because the company has engineered them to point the same way. That's a strength when it produces models that say "no" to garbage, and a liability when "unsafe to release" becomes indistinguishable from "ours to control." The thread's most honest position isn't pro or anti — it's that a single firm's sincere belief is now a governance input for a technology touching billions, and sincerity is not the same thing as accountability.