BEFORE: Opus 5.5 does everything. Reads every file. Decides every retry. $6.42 per task. AFTER: Jev makes the small decisions. Opus only does the hard reasoning. Confidence thresholds catch the risky calls. $2.00 per task. Same agent. Same code quality. A third of the bill. Full build below 👇

the confidence threshold idea is solid but curious what your actual false positive rate looks like on those easy calls before you filter to opus

$6.42 down to $2.00 just by letting opus 5.5 skip the small stuff, smart

did the confidence thresholds need calibration per-task or did you find one value that worked across the board? 🎯

This is where model routing delivers real savings.

A good leader outsources thinking that can and should be done by the team to the team. Works for models too.

i think the catch is the calls the threshold sends up, opus now gets the hard ones with less context than when it read everything itself

Routing the small decisions to a cheaper model is the real cost win. Most retry loops do not need the big model every time.

i used to let the main model decide every tiny retry too. that gets expensive way faster than you expect