Claude Code found $710/mo hiding inside an agent loop that already looked “optimized” /cost-audit 1. feed it the usage log for every task 2. make it flag late Opus escalations, cold-cache rewrites and effort changes 3. let it rewrite the router, handoff and session rules around those failures this... run scanned 4,812 calls across 1,000 tasks and found 3 leaks the biggest was late escalation from Sonnet to Opus after the fixes: $2,530/mo → $1,820/mo $710 saved on the same workload Claude Code can audit the system, not just write the code ↓show more

I won't make such mistakes anymore

the craziest part isn't saving $710/mo it's finding that $710 inside a loop everyone already thought was optimized

first, detection, awareness, and cost savings congrats

late escalation being the top leak tracks with what i see

that’s right it is happening exactly as you see it

nice work, that cost-audit setup is clean

the cost audit setup is one of the cleanest I’ve ever seen

best command

The retries are always where the money hides imo, not the model price

nice found

It actually makes so much senses

everything is easy on the inside, but from the outside, it all looks difficult

thank you for this useful content Gipp

glad to hear it, brother

finding $710 is quite a challenge. claude managed to do it.

yes claude is a real genius

nice work

thank my homie

gipp, thanks for sharing this honestly never even thought about it

I would like to use this too

Do the new router rules add noticeable latency for real-time calls?

dynamic effort adjustments prevent paying top rates on easy tasks

good idea how to save credits on unimportant tasks

yep yep

this looks super interesting gipp

I think there is room for further work here

Cold-cache rewrites are the sneaky one. Which of the three saved you the most? @claudeai

Late escalation is the sneaky one. By the time Sonnet hands off to Opus the context is already huge, so you pay Opus rates on the whole mess plus a cold cache. Picking the model by task type up front is usually cheaper than escalating mid-loop.

cool

did you look at how the bill changes if you escalate earlier?

finding hidden costs like that can really boost efficiency and ROI

Claude catching its own Opus habit is the funniest part. Those leaks hide until you actually log everything.

I didn't know that command even existed thanks, bro, for finding it

interesting find

maybe i'll find some unused money that i spend monthly

This one is going in my greatest hits folder

the fixes matter

Pay attention to it

Late escalation costs twice: you pay for the weaker model's failed attempts, then the stronger model rebuilds the whole cache. Routing on task shape before the first call, and handing over a compact summary instead of raw context when you do switch, cuts both.

bookmark this alpha

Auditing the agent itself is a powerful idea.

late escalations and cold cache rewrites are where most agent bills leak. routing on task type upfront beats escalating to the big model at step 9