Realistically they can't go that much above the actual cost for inference since the customer can always switch to self hosted or inference only providers. Or their models have to be significantly better than the open source models for the foreseeable future. They will never be able to charge much more for their lower tier models.