“The per-token inference cost for frontier AI models may not increase, as algorithmic efficiencies could offset the costs of larger models.”