Does it really decrease token use by 95%

[princecapitalgrp]

Im curious because google, Claude Code & instagram comments all say that its likely not doing that but how provable is this?

[jaramkt]

I've been a little confused about this topic, because since I install it, my usage has been going up really fast like never before!

[octavianbudai]

Same here, I installed it two days ago and my token usage sky rocketed. Two basic prompts and i'm at 100% session. Dunno what's causing it but I'm tired of it. Uninstalling asap.

[274188A]

same - i don't for a sec now believe the

[md-mashiur]

don't ever install it again
https://prnt.sc/4Pv_8Z2t2JFT

[octavianbudai]

I uninstalled it yesterday and everything is fine now. Token usage is back to normal.

[rrajan94]

#2128
I guess because of this token usage going up.

[octavianbudai]

Yep and I couldnt figure out why. Seems like a catastrophical bug to me. I uninstalled and everything is fine now.

[apprentice10]

Same here, I think It would make more sense if instead of using the same claude account assing a local AI to do that job in the background

[maxencemeloni]

same for me, uninstalled, i didnt noticed, but 1 week of usage, it consumes more than my project itself.

[whoishack]

yeah this is bogus.
i've had my computer off. booted up, opened CC and asked 1 question, used 20%. went to kitchen and got coffee and came back and suddenly claude reports usage at 90%. look at all these background claude-mem sessions! I even had all terminals closed and came on here to see what others are saying, now claude reports 100% used.

[sean3808]

My perspective is that we shouldn't just look at the token count; we need to focus on the overall price.

I have two main points regarding the cost:

1. The system defaults to using Sonnet. If we switch to Haiku instead, the costs will decrease significantly.
2. While their advertising claims a 95% reduction in costs for retrieval tasks, they fail to mention that building the necessary infrastructure involves separate costs. So, while the 95% savings might be accurate for the retrieval process itself, the setup costs are an entirely different matter.

To lower the costs for the latter, we have two options: either switch from Sonnet to Haiku, or integrate a local LLM to handle that portion of the workload.