The safety tax argument
A newsletter revives OpenAI's case that teaching a model to reason over its own safety rules makes it both harder to jailbreak and less prone to refusing harmless questions — the optimistic half of a debate whose other half runs the opposite way.
The claim being resurfaced is concrete and checkable. Instead of training a model on labelled examples of good and bad behaviour, OpenAI gave o1 the actual text of its safety policy and taught it to reason over those rules in its own working-out before answering. On a standard jailbreak test the reasoning model scored 0.88 where the previous flagship, GPT-4o, managed 0.37 — and it did this while turning away fewer perfectly benign questions, not more. Getting harder to trick usually makes a model more twitchy about innocent prompts; here both numbers moved the right way at once.
That result is real, and it is old. It comes from a paper published in December 2024 and baked into shipped models since — not a 2026 development. What is new is the framing wrapped around it: a newsletter cites it, without a citation, to argue that smarter AI will therefore be safer AI, capability and alignment climbing together as a general law.
The paper does not say that. It shows one training method improved one tradeoff — a narrow, honest result. And a parallel research line, published under the deadpan name 'Safety Tax,' finds the reverse: bolting safety alignment onto a reasoning model can measurably degrade its reasoning. The same recent literature the optimism draws on also contains a well-cited body arguing capability and safety trade off against each other. The eighteen-month-old benchmark is not the story; both halves of that argument are live at once, and the cheerful version only ever quotes one of them.
The lenses
The facts
How this connects
Tap a node to open it