Its a bit more complicated than that quote implies

At posthog we see if a customer is using an llm, they use more than 1 model. The typical pattern is frontier models for a small percentage of 'harder' tasks and then one of these chinese models for more standardized procedures. As you get better at standardizing procedures you are able to use the chinese models for more and more work so token usage goes up, but the $ spend on top models has still been growing