← all posts site ↗

the model is the price, not the harness

kept circling a question: is another coding agent harness cheaper than the one i use, at the same model? the instinct was that a leaner tool would slash my bills. published head to head data actually exists, thirty real tasks, same machine, same model across three top agents: the one i use burns roughly one and a half to two times the tokens of the leaner ones per task, and lands around double the cost. so yes, the harness matters, mostly because it drags a huge system prompt around every turn and reads whole files where a leaner tool reads just the symbols that changed.

but the real killer in my own history was never tokens, it was the model. one week on a fancier model cost more than a nice dinner out, and something like ninety percent of the volume was cache reads, billed at a per token price many times higher than the cheap model i normally pin. the model lever was a twelve times swing. the harness lever was maybe two to three times.

the conclusion landed hard: the model you pick can out destroy any harness inefficiency. GLM on a fat context costs more than the cheap model ever will, no matter which agent drives it. the habits that stuck: pin the cheap model for all the app coding, start a fresh session at each feature checkpoint because context growth is the real tax, and only then is a leaner harness a bonus on top, not the fix. the model is the price tag, the harness just multiplies what you feed it.