Activity Monitor sucks, alternatives were too expensive and not that good, so I started building a AGPL replacement: https://t.co/GqRQ8AaGyh
0.54 tonight baby. With now processus typing, ring pinning, and a shitload of cleanup :D
@RafikSmati L'enjeu est de maîtriser les deux. Parce que racheter un modèle européen pour un acteur étranger ça sera plus simple que de déplacer des GPUs.
Donc l'infra ça compte. Pour le reste, le problème est européen. Il suffit de dupliquer le capitalisme d'état US et la donne est changée
Hey @ClaudeDevs Opus 4.8 way of spawning Haiku subagents consume up to 6 times more token.
Since May 29th, Opus use a new template to prompt subagents exploration that is leading to a huge increase in token consumptions. My guess would be it's due to the verbatim code order.
@ClaudeDevs Found it! Before May 29th, Opus used Subagents to map/audit targeted things since yesterday Opus leverages subagents very differently : previous plan is shipped > different task > overwrite plan > explore/map N domains in parallel > each agent report file:line + verbatim code
@ClaudeDevs Found it! Before May 29th, Opus used Subagents to map/audit targeted things since yesterday Opus leverages subagents very differently : previous plan is shipped > different task > overwrite plan > explore/map N domains in parallel > each agent report file:line + verbatim code
At least data is very clear now, the 20x subscription is equal to 25 hours of work with Opus 4.8 at maximum session usage.
To reach that I had 4 real sessions running at the same time. It was a long time since I maxed out a session. I'm trying to find the culprit.
@ClaudeDevs and to give you some context on May 29 I've had a major work done on a quite large Django backend.
If we'd calculate a ratio, until May 29 we were at a 10k/agent ratio, 29th 15k and 30th 20k. Probably the change happened on the 29th. Let's dive.
And that's twice the cache in 30% of the work time at maximum. So there definitely is a cache thing going, and caches are cheaper but they do count into quotas and at that usage rate (6x from yesterday) no surprise my session filled up fast.