What counts as cheaper AI inference?
One participant says the cost of using a given AI model falls by roughly 30x a year. Another argues that most people aren't running older models and cites a price increase from 4o to Astra.
TLDR
The lighting analogy frames the exchange: one user claims lighting costs have fallen by over 1,000x, and another predicts the same for AI inference. The disagreement centers on costs for a given model versus the price of newer models. The first user puts the decline for a given model at “something like 30x a year”; the other says demand for older models is low and that the price rose from 4o to Astra. Invoking Jevons Paradox, the first user argues that lower unit prices lead people to use more—and says the relevant unit isn't the token.
Combined views
91.9K
6 Sources, first seen 16d ago