From your link:

"We excluded reasoning models from our analysis of per-token prices. Reasoning models tend to generate a much larger number of tokens than other models, making these models cost more in total to evaluate on a benchmark. This makes it misleading to compare reasoning models to other models on price per token, at a given performance level."

It's just price per token. Token usage is exploding.

Token usage is increasing but the link I shared is comparing price per token which is the reason they are excluding reasoning models.

Reasoning models have also gotten cheaper.

Generally cost to “achieve a certain task” using whatever model, even reasoning has drastically reduced.

The best example is arc agi. https://arcprize.org/leaderboard

This measures cost to achieve certain percentage of score. Fix a certain accuracy and see how much price reduces over time. It’s more than 100x.