A rate card prices tokens. A reasoning model decides for itself how many tokens your task needs, and you learn that number after it has been generated. That is the whole distance between 16x on paper [3] and 3.4x on the bill [2].
Divide tokens by seconds and the speed claim gets more interesting. On the date parser, Flash produced 38,677 tokens in 455.8 seconds [7] and the flagship 14,801 in 174.7 seconds [8], which is 84.9 against 84.7 tokens per second [4]. Flash's additional latency comes from generating 2.6x more tokens at roughly the same rate [1]. The advertised 3x serving improvement [1] does not show up anywhere in that wall clock. Treat both rates as approximate: the 38,677 figure is described as reasoning tokens with a 60-line function on top of it [7], and a hosted endpoint puts queueing inside your latency.
The puzzle runs the other way. Flash answered with 1,003 output tokens in 33.5 seconds, the flagship with 1,804 in 18.9 [13], which works out to 29.9 tokens per second against 95.5 [5]. Same two models, same test harness, and the throughput ordering flips.
The cost ordering flips in size too. Priced at the posted output rates, the flagship's scheduling answer costs $0.0075 and Flash's costs $0.00025, a 30x gap [6]. On the date parser the realized gap was 3.4x [2]. That is a spread of nearly 9x between two tasks [7] on one pair of endpoints. Cost per million tokens never moves. Cost per completed task swings by nearly 9x between these two tasks, and that number is the one that lands on your monthly bill.
One number here does not reconcile. Apply the flagship's $4.18 per million output tokens to its 14,801 tokens and you get $0.062, close to the $0.065 reported. Apply Flash's $0.25 to its 38,677 tokens and you get $0.0097, about half the $0.019 reported [8]. So either more tokens were billed than were reported, or the account was not charged the listed rate. Take the posted rates literally for both and the date-parser gap narrows to 6.4x [9], which is not 16 either.
For any of this to transfer, your prompts would have to pull a similar reasoning length out of Flash, and your acceptance criteria would have to look like a 12-case hidden suite that both models passed clean [6]. The grading is better than most benchmark tables offer: the suite was written and verified before either model saw the spec [5], and all 120 candidate schedules were brute-forced in advance to confirm exactly one solution existed [12]. That is the part worth copying. Spending 38,677 tokens to arrive at a 60-line date parser is a lot of deliberation for a function the standard library nearly ships.
The New Stack's own read is that the budget model's low price covers for it working harder to reach the same answer [11], and that a pricing change at Z.AI could invert the arithmetic entirely [14]. On this evidence the flagship's premium buys shorter reasoning, and whether that is cheaper depends on a token count you can only measure after the run.
