Imagine a procurement team comparing two AI tiers. The "reasoning" one costs several times more per query, thinks for thirty seconds, and produces a cleaner answer on a benchmark. The vendor calls that a breakthrough. But what changed is the pricing menu and the compute bill, not necessarily the underlying capability. The incentive is obvious: if buyers pay a premium for a badge, labs will slap "reasoning" on everything and optimize for the tests buyers see. That's Goodhart's law with a credit card. The other side will say these models solve harder problems. Sometimes, yes. But a breakthrough should show broad, reliable gains at sensible cost. Instead we get benchmark gains bought with huge token budgets, and users pay for the search. That's a marketing equilibrium, not a scientific one.

Let's frame this as a production problem. If the only consistent way to get the better answer is to buy far more inference tokens, that's not a new capability, it's a more expensive input. My opponent's scheduling example is fair, but it proves the point: the model gets better because someone paid for a minute of search. Then the market labels that premium tier reasoning and everyone copies it. A real breakthrough would lower cost per reliable decision or deliver gains that transfer broadly without the token burn. Benchmark wins purchased with compute are a pricing strategy, and buyers should judge the invoice, not the badge.

Does a 74 percent score still count as a breakthrough if the invoice is ten times higher? My opponent's AIME example is real, but it actually shows the mechanism: more inference tokens buy more search. That's a costly input, not proof of a new capability. And the premium can survive because buyers face information asymmetry and switching costs, not because repeated play cleans it up. So judge cost per reliable decision, not the badge or a benchmark bought with token burn.