The biggest problem with this is that I can't run my company's own internal benchmarks against this API, because they will be trained into the next version of the model, making them worthless.