The most interesting thing about this round of cost cuts by homegrown models isn’t that they’re cheaper—it’s that they’re forcing everyone to run the numbers again.

People used to compete over “who’s smarter.” Now they’re starting to compete over “who can do the same thing for less.”

This shift is actually good for ordinary users: you shouldn’t have to pay for model capabilities you don’t need.

What’s really worth watching is which products that used to be uneconomical suddenly become viable once inference costs fall below a certain threshold.

The cost curve determines the shape of the product, not the other way around.