Reasoning models, sometimes called thinking models, work through a problem step by step before answering. That intermediate work is produced as tokens. Providers bill it as output, the most expensive kind of token, whether or not the user ever sees it.
For a short factual question the reasoning may be small. For a hard analysis, a coding task or an agent planning its next step, the reasoning can run to several times the length of the visible answer. Because output is priced at several times input, the effect on the bill is larger still.
Three practical consequences follow for finance and FinOps teams:
- Measure it. Most providers report reasoning tokens separately in the usage data returned with each call. Cost dashboards that show only total output hide the reasoning share, so read the usage fields and track the share per workload.
- Budget it. Many models let the caller set a reasoning effort or a thinking budget. Set it per workload: deep reasoning where accuracy pays for it, light or none where it does not.
- Route around it. Tasks that do not need deep reasoning can go to models or settings that do not use it. As with any routing decision, check that the cheaper route still clears the quality floor.
Reasoning is one reason the newest frontier models have not fallen in price as fast as older capability levels, and one reason AI bills keep rising while list prices fall. See Token Economics for the other mechanics of a token bill.