“The model referred to as "GPT 5.4" is significantly cheaper to serve than GPT-4 because it has fewer active parameters and utilizes a sparser Mixture of Experts (MoE) architecture.”