In the 2022 paper introducing OPT (Open Pre-trained Transformer), Meta AI stated that OPT-175B was comparable to GPT-3 while requiring only one-seventh of the carbon footprint to develop. Both models held 175 billion parameters, so the claim was about the efficiency of producing a model of that scale, not a difference in size. The figure became a frequently cited data point in discussions of the energy and emissions cost of training very large language models.