the per-token price does not move when a model thinks harder. what changes is the budget. the reasoning setting controls the share of output tokens spent thinking, roughly a tenth at minimal, a fifth at low, half at medium, four fifths at high, and about nineteen twentieths at max per the provider's mapping. a max request can bill for several times the output of a low request, because thinking is billed at the output rate.
the practical answer is that it rarely matters on the cheap default model. the absolute dollars per message are fractions of a cent either way. it matters on chains that run thousands of calls, where dropping from max to high saves about fifteen percent of thinking tokens on the tasks that actually use the full budget.
so the right default is high, and max only when a visible task fails there. the quality gap shows up mostly on hard reasoning, and you get the latency back, which is worth more than the billed tokens in an agent loop.