I benchmarked Claude Code six ways — Sonnet plus a 4-bullet prompt won
Six controlled runs, one engineering ticket, configurations A through F. Sonnet 4.6 plus a four-bullet efficiency addendum used 27% fewer turns and 24% fewer output tokens than the same model unprompted — and the more expensive Opus 4.7 brought zero quality gain at 1.5–2× the cost.
Read article →