The price behind the price The difference between a price tag and an invoice

A score tells you how well a model responds. The price it charges for that doesn't yet reveal how expensive it will be to run. CrucibleMark separates the two: actual token consumption across all benchmark tasks and the resulting total cost per run. What becomes visible is a cost range that often has little to do with the listed rate.

Token consumption reveals what a score keeps quiet. That sounds obvious, but it still caught me off guard in real-world operation. In direct API use — say, coding with an assistant — output drives the price: how much a model reasons and how verbosely it responds. Anyone using a chatty model as a coding assistant notices this at the latest when the next invoice arrives. In an agent workflow, the dynamic flips. A model is no longer answering individual questions; it works for hours, with a growing context on every request. My rule of thumb: roughly 98 percent of the tokens an agent consumes are prompt tokens. Not answers — memory. What looks harmless at ten requests becomes a cost trap at 300. The same principle applies to both scenarios: the rate beats verbosity. Anyone running an expensive model in an agent workflow pays that rate not once, but with every request, for the same growing context.

The chart below shows what a complete benchmark run costs: measured against actual token consumption and each model's rate. Sortable by score, consumption, or cost — three perspectives on the same models.