Comment by marcinjachymiak

It's not surprising since these are all reasoning models now. Token sampling/decoding isn't exposed because the reasoning models are trained around specific configurations. That's why it's been replaced with "reasoning effort" instead (low, medium, high "thinking" levels).