My reading on this is that they refuse to allow public access to newer models because inference is much more expensive than the $20 per month that the market is

My reading on this is that they refuse to allow public access to newer models because inference is much more expensive than the $20 per month that the market is willing to pay. That's why they are working on dedicated inference hardware. And that's why they are still losing money: they are selling access to their AI for less than what it costs to operate. Releasing more resource intensive products (without raising prices) would then just increase their losses. And investor money being limited then forces them to not release new products.