The switch by Artificial Analysis from per-token-cost to per-benchmark-cost shows some effect! Its nice that labs are now trying to optimize what I actually have to pay to get an answer - It always annoys me to have to pay for all the senseless rambling of the less-capable reasoning models.