Yes, skills that are actually used count toward token consumption.

The question is whether the number of tokens required to achieve a certain behavior/intelligence/quality is equal between you manually providing those tokens via skills versus the model "deriving" the "skills" it needs on-the-fly in order to produce the outcome you want.

The claim above is that the former requires far fewer tokens.

Also skills only consume tokens when they are used, and part of the value is that the model will dynamically find and disclose only what's needed (assuming the skill is "well-designed").

Their claim is not about the prompt or skill tokens, it's about output tokens - skills can help the model bypass some thinking tokens or avoid reasoning deadends, and that way reduce output token usage. That's what they seem to have found empirically from their testing. (If it's truly 2x-4x, the time savings in waiting for the output is a pretty nice benefit too.)