You're not running an Opus 4.8/GPT 5.6 tier model on $5k of hardware at a useful Tok/s