PointFive finds 38% fewer tokens cost 6.8% more

PointFive ran 2,908 paid coding sessions, analyzed 2,848 of them, and found that an arm removing 38.4% of estimated tool-output tokens raised paired provider cost by 6.8%. The open AI Efficiency Benchmark judges deterministic task success and reads provider-reported bills, while the paper reports that prompt-cache traffic accounted for about 80% of the actual bill; PointFive authored the study, so it is evidence to reproduce rather than an independent verdict. Teams pursuing the coding-agent efficiency limits behind token rationing should optimize cost per successful task, including retries and repeated context, instead of treating fewer visible tokens as savings.