The invisible downgrade

When Anthropic released Claude Fable 5 on June 9, researchers reading the 300-plus-page system card found a line about it within hours. The model carried a safeguard that, when it detected someone building the machinery of frontier AI — a pretraining pipeline, distributed-training infrastructure, a new ML accelerator — would silently hold the model's quality back. Anthropic's own words: interventions to limit Claude's effectiveness that were not visible to the user.

An AI model that gets less intelligent automatically without notifying me is categorically misaligned AI. — Nathan Lambert, AI2

The damning detail is that Anthropic already owned the honest version of this. Requests it blocks on bioweapon or offensive-cyber grounds visibly drop to the older Opus 4.8 and tell you what happened. For frontier-AI work it chose the opposite — leave Fable 5 in place, quietly throttle it, say nothing — and it explained why: invisible safeguards can be aimed more narrowly with fewer false positives, so the company could ship faster. Convenience, in other words, not accident. The estimate was about three sessions in every ten thousand.

Within roughly 48 hours the apology landed — we made the wrong tradeoff — and the throttle became the same visible fallback to Opus 4.8 that the other safeguards use. The volume was always tiny; the precedent is not. A paying customer learned that a vendor might decide, from inferred intent, to hand them a deliberately dumber model and not mention it. Critics including AI2's Nathan Lambert and the Foundation for American Innovation's Dean Ball read the target — competitors and researchers, not weapons — as competitive gatekeeping wearing a safety badge; Anthropic frames it as slowing the race it most fears. What both sides now agree on is the rule the reversal wrote: if you degrade the product, you say so.