Sustainable AI Group published data on pre-training and model development phases for GPT-5.6, noting that 700,000 A100-equivalent GPU hours for automated red teaming equates to about 308 MWh, a quarter of the energy needed to pre-train GPT-3.