There are two model leaderboards on the same page, and they barely overlap

Strip the two lists down and you are looking at the difference between what the industry talks about and what it buys. The conversation is about frontier capability. The traffic is overwhelmingly cheap, fast and, increasingly, Chinese open-weight. That is not developers settling for less. It is thousands of teams having already run the experiment most organisations are still holding workshops about: route the routine eighty per cent to something cheap that is good enough, and keep the expensive model for the fifth of the work that genuinely needs it. The word doing the heavy lifting in that usage list is "Flash". Three of the top six are explicitly the budget tier of a capable family. Nobody chose those because they topped a benchmark. They chose them because an agent that re-reads the same context on every call makes price per million the number that decides whether the product works commercially. Two things follow if you are building. First, if you are running everything through one frontier model because it scored well, you are paying a premium the market has already stopped paying, and there is now a year of public evidence about where the cheap tier does and doesn't hold up. Second — and this is the less comfortable one — the benchmark you are quoting in your deck is measuring something other than what your users will experience. The revealed preference is a better guide than the leaderboard, and it points at a tier most decks don't mention. Worth naming what this doesn't say. It doesn't say the cheap models are as good; volume is not quality, and Anthropic's share of dollars runs far ahead of its share of tokens for a reason.