The distinction between a within-model language premium and a cross-model price comparison is especially useful here. For closed-model measurements, I would also retain the exact submitted messages and reported usage beside each text pair. Subtracting a prompt-only request assumes the gateway adds the same fixed overhead in both cases, which is worth checking rather than treating as invisible.

The English-instructions recommendation could be tested separately from translating user content. Keeping the original support message while translating only stable instructions and tool descriptions would show the achievable token reduction without losing names, quoted passages, or language-specific intent through an extra translation step.