Luckily the models were built on copyrighted materials so hopefully the big players won't have strong legal standing to kill off model destilling. Then, with models like current deepseek or kimi k2.5, you are perhapy 1/2 years behind in their capabilities at the fraction of the cost. For real inference costs, look at prices on openrouter. For hobby, I wasn't able to burn trough more than $5 in a month.