Inference costs can't be faked though, since the model can be run locally by anyone with capable hardware.

Even if the whole story about the training cost was fake, R1 and the distilled models are still very efficient at inference.

The shock for the industry was the claimed training costs and used hardware.

unless OAI has all the optimization tricks already, which they probably do

Is the model architecture actually that different from anything else? Or are you just saying that you can get away with smaller models now?

What DS did is in line with my expectations as I see a lot of performance optimizations possible at algorithmic level. So, even if the DS numbers are BS somebody else will tomorrow reach and even beat it.