[svachalek]

Training frontier models is expensive. Running inference on them is pretty cheap and for solving the same level of problem, will continue to get cheaper. The more use they get out of a model, the less overhead we need to pay on that inference to subsidize the training.