[bostik]

I can think of two reasons. One, to have the plausible deniability with the necessary future statement "Claude is not used by the DoD/DoW to conduct domestic mass surveillance or autonomous killing"; by having the model be properly a different from the one used by the public, they can wrangle over the language with technicalities and still avoid outright lying. (With their IPO in sight, let's keep in mind that everything is securities fraud.)

And two, I suspect that some of the guardrails have been "baked in" to Anthropic's model. Much in the same way as the Chinese open-weight models have a strong bias against expressing positive sentiments about Tiananmen Square, Tank Man or Winnie the Pooh, the "Standard Claude" would likely have the fundamental product biases trained into it.

Taken together it would therefore be both politically and financially sensible for Anthropic to create a separate, unrestricted[tm] almost-Claude for the morally unconstrained military / intelligence purposes.