Per-token prices are flat or falling. Cost per task is going the other way.

Cost per task is the product of two numbers, and buyers only negotiate one of them. The price sheet is per token; the invoice is per token consumed. Between Grok 4.5 (high) and Grok 4.6 (high), SpaceXAI changed neither side of its list price ^[3][7], and the Intelligence Index run still cost 84 percent more ^[1]. Dollars per index point went from $10.34 to $17.52, a 69 percent rise ^[2]. Alibaba is the harder case for the deflation story, because Alibaba actually cut prices ^[11][12], and the benchmark bill rose 64 percent anyway ^[3], with cost per point up a third ^[4]. Qwen3.8 Max generated 150M output tokens on the run against Qwen3.7 Max's 100M ^[10][12]; half again the output at four fifths the price should have produced a 20 percent increase, not 64 ^[7].

There is a reconciliation problem worth flagging in the Grok pair. Output volume went from 60M tokens to 72M ^[2][6], which at unchanged rates implies a bill about a fifth larger ^[6]. Either those token counts are rounded past the point of usefulness, or the growth sits on the input side, context re-sent across longer runs. Time to first answer token moved from 5.63s to 48.42s ^[8][4], consistent with considerably more work per question.

Meta is candid about the one lever still moving fast. Its contributor tier is 12.5 times cheaper on input and 21.25 times cheaper on output than the standard tier ^[19][20], which would take Muse Spark 1.2 from $0.40 per task to roughly two or three cents ^[8]. That deflation is real. It is settled in code, not dollars ^[20].