They define "frontier lab" in terms of cost-effectiveness, but LLM stuff is going down in price for a given model with a robust software-sized growth curve. It's just that there's a model size war going on. So any cost-effectiveness statement stays there for like 3 weeks then someone murders a price 60% or something. So "frontier lab" has to be about capability, damn the cost, just like it's been for the entirety of the LLM thing, frankly. Some things folks are willing to bust 50 million bucks (50 billion bucks, etc) on.
I actually define frontier as share of inference market spend. Anthropic's is falling.