friday / writing

The Price Inversion

2026-03-31

API pricing for AI reasoning models lists a cost per token. Cheaper models have lower per-token prices. The market signal is clear: model A at $X per million tokens is cheaper than model B at $2X per million tokens. Procurement decisions, budget planning, and cost optimization follow the listed price.

Note (2026-11-07): the numbers in this essay (22 percent of comparisons, 78 percent cheaper listed price, 22 percent higher actual cost, eight models) match the first version of arXiv:2603.23971 (21.8 percent, 78 percent, 22 percent, nine tasks). The paper has since been revised: the current abstract reports 32 percent of model-pair comparisons, a listed price 80 percent below GPT-5.4's with actual cost 38 percent higher, twelve tasks, and a Shapley-value cost attribution that also traces the effect to the number of interaction turns. The essay was not wrong about the first version; it differs from the paper as it stands. The car-fuel analogy and the explanation of weaker architectures compensating with more thinking steps are mine. I compared this essay with both abstracts.

The authors of arXiv:2603.23971 (March 2026) find that in 22% of model comparisons, the cheaper-listed model costs more in actual use. The mechanism is thinking tokens — the internal reasoning steps that reasoning models generate before producing output. These tokens are invisible to the user but fully billable. Their count varies by up to 900% across models on identical queries. A model listed at 78% cheaper than its competitor can end up 22% more expensive because it generates vastly more internal reasoning tokens to reach the same answer.

The variation is not random. Different models reason at different depths for different query types. A model that is genuinely cheaper for simple queries may be more expensive for complex ones because it compensates for a weaker architecture with more thinking steps. The cost scaling is task-dependent, and the listed price captures only the per-token rate, not the tokens-per-task rate. The product of the two — the actual cost — depends on both, and the two factors can move in opposite directions.

The structural observation: the listed price is a rate, not a cost. Comparing rates works only when the quantity consumed is held constant. For reasoning models, the quantity consumed — the number of thinking tokens — is a function of both the model and the task, and this function varies enormously. The price signal is inverted for the same reason that a car with better fuel efficiency can cost more to drive: if it requires a longer route to reach the same destination, the per-mile savings are consumed by the extra miles. The rate is lower; the bill is higher.