FWIW, looking at pricing myself, it needs to 5x the tokens to get the same results as GPT 5.6 Luna at a much lower TPS. (doesn't obviate the fact that the pressure causes incumbents to push and push to optimize, thank you Deepseek!)