There is simply not adequate proof of this for Anthropic, certainly not for OpenAI. All of their figures are distorted by ignoring the costs of training, free tier usage; even seem to ignore the spent cost of training the current models, let alone the future models.