Ben Thompson, the tech industry analyst who publishes the Stratechery newsletter, reads Anthropic's pattern of previewing a restricted AI model and then shipping a public version with guardrails as a strategic advantage, and takes the cynic's view seriously.
Anthropic's pattern of previewing a frontier AI model as too dangerous to release, then shipping a guarded public version roughly two months later, is either a meaningful safety practice or a marketing maneuver. Stratechery analyst Ben Thompson argues it is closer to the former, and that the framing itself has become a strategic asset, what he calls a Safety Superpower.
The argument runs through a specific sequence of moves. Roughly two months before the public release of a model called Fable, Anthropic announced a preview of a different model, Mythos. The company publicly characterized Mythos as too dangerous to release, citing advanced cybersecurity capabilities. About two months later, Anthropic released Fable, framed as Mythos with safety guardrails applied. The pattern is the point: a restricted preview establishes the ceiling of what the underlying model can do, and the public release positions the guardrails as the differentiator. (Anthropic's Safety Superpower)
Thompson's read of the capability story is subjective but confident. He reports that Fable made current competing models feel "small and dumb" in direct interaction, and he places Fable in the same tier as two earlier releases he had previously associated with new generations in base model size and complexity (Thompson names GPT-4 and Grok 4 as the prior reference points). His working theory: Fable is downstream of a new pre-train and is the first model of a new generation. (Anthropic's Safety Superpower)
The strength of those capability claims is what gives the safety framing weight. If Fable were merely competitive with peers, the "too dangerous to release" preview would read as marketing. If Fable is genuinely at the frontier, the same framing reads as a legitimate disclosure pattern, the kind of move a safety-conscious lab is supposed to make.
Thompson takes the cynic's view seriously and does not dismiss it. The read that Anthropic's safety framing is a marketing layer over a normal product cycle is coherent, and the company's commercial incentives cut in that direction. He complicates that view by pointing to the substantive capability gains, and he is honest that the two readings, marketing and genuine safety leadership, are not mutually exclusive. (Anthropic's Safety Superpower)
A few caveats apply. Stratechery is an opinion and analysis publication, and Thompson's model evaluation is one analyst's subjective impression, not an industry consensus. The "Mythos too dangerous" framing originates with Anthropic and reaches readers through Thompson; primary Anthropic announcement URLs or independent confirmation of the specific capability claims would strengthen the case. The model comparisons in the piece are Thompson's read, not benchmark results.
The open question, then, is whether the Safety Superpower framing survives contact with the next release cycle. If Anthropic continues to preview restricted models and ship guarded public versions, and if the underlying capability keeps moving, the framing hardens into an industry expectation. If the pattern breaks, or if a peer lab ships a comparable model without the safety preamble, the cynic's view hardens instead. Either way, the gap between restricted preview and public release is now a strategic variable other frontier labs will have to respond to. (Anthropic's Safety Superpower)