Six502 @_Six502

Week one of selling my agent-cost tool, in numbers. LinkedIn post: 15,000 impressions. Free router: 79 clones. Product site: 221 visits, 0 signups. Best post on here: 29 views. Hi :) So people LOVE reading about the bill. Fixing it? Not yet.

Run enough subagents in one repo and two will edit the same file. Last write wins, and nobody tells you. So my swarm orchestrator now lets the first one hold the file (10 mins). Anyone else gets "X is editing it, message them or wait". Blind spot? sed -i sails straight past.

@hayatomaruu Also , Is Jev all its cracked up to be? I measured it : fortitude-omnis.group/rd/tau/

@hayatomaruu I must be in the 1% :), I tried RuFlo, bit too busy, leaky and slow. So I wrote my own swarm orchestrator. The little claudes dance and their arms wag when they make tool calls .

Autocompaction is now in 2.1.299 so this is now moot

Got asked if you could ping Claude Code every 59 minutes to keep the 1-hour cache warm. Did the sums on my logs. A warm read of a ~470K session: about $0.24. Rewriting it after the cache dies: $4.70. So pinging wins until you've been gone ~20 hours. Lunch, yes. Weekend, NO.

@SteamedBun1111 @goon_nguyen Yep, that's it in one line. The threshold sets how big each rewrite is, walking away sets how often. Idle compaction fixes the second, if you're happy coming back to a summary rather than the real thing. I think its designed to make Anthropic richer 😂

207 times in 33 days I came back to a Claude Code session after the cache had expired. Each time it rewrote ~470k tokens of context. $972 at API rates. 22% of my main-session spend, for walking away from the keyboard! Where the rest of the bill goes: omnisvigil.fortitude-omnis.group/articles/ai-co…

@SteamedBun1111 @goon_nguyen Honestly? No, still on the default. When I dug further, the bigger leak was the gaps: 207 times I came back after the 1h cache expired and rewrote ~470k tokens, $972 of it. A lower threshold shrinks the reads, but every compaction is a fresh cache write too. Nice slider!

@jacksunny18 You're right, I checked. Confidence there is an entropy (or margin) figure, not a calibrated probability, and the README example prints it with no warning. What you gate on is the chosen answer's probability after calibration. Fixing the README, thanks for spotting it!

I was paying Claude Opus to tell me "my card payment failed" is a billing question. Every time. So I stuck my gaming GPU in front of it. 73% of the decisions never left the card, for about a quarter of the bill. Then a 22M-parameter classifier beat the lot :/

@itsvlady Flat subscription, so the real bill's boring. But I priced 33 days of my Claude Code logs at API rates and got ~$4,560, 92% of it Opus. Only 8% of that was the model writing anything, the rest was context and cache. Which is why I'm never giving up the plan!

@Skaly__Bull Checked mine against that line. Over 33 days, 54% of my subagent requests in Claude Code were over 100K tokens (mostly cached context they reread). So if I swapped them all to Haiku 5.5, half would land on the $0.50/$2.50 card. Cheap model, wrong side of the cliff :/

@lydiahallie Checked what this was costing me. 33 days of my Claude Code logs: 207 times I came back after the hour, and each one rewrote ~470k tokens of context to cache. $972 at API rates, 22% of my main-session spend. The same context read warm would've been $48. Good change!

@Arindam_1729 My logs agree. 33 days of Claude Code, 92% of the spend on Opus. But output was only 8% of the bill and cache reads 62%, so the model you pick mostly sets the price of rereading context. Haiku doing the read-heavy exploring is where a split like this really pays.

@rryssf Measured this before flipping it. 33 days of my own Claude Code logs: subagents only ever wrote 5-minute cache, and ALL subagent traffic was 4% of the bill ($177 of ~$4,560 at API rates). The main session was the other 96%, and it already writes the 1h cache. Check yours first!

@mariioalexandre Sorry for the delay, I was away on holiday in Cornwall! Good call though. Free-form summaries are where my subagents get chatty, and thesession pays for every extra line again on each later turn. Going to try a fixed status/files/findings shape this month & see if the reads drop.

Where does an A.I. agent's money go? I checked 33 days of my own logs. The code it writes? 8% of the bill. Rereading stuff it's ALREADY seen? 62%! Turns out the expensive bit is the remembering. Who knew? Keep sessions short, make subagents hand back a summary & bin the rest.

@Ramabdel12 @skunzu Yeh, you're right on the raw API. The advisor tool takes a caching option (ephemeral, 5m or 1h) and breaks even at about 3 calls a conversation. My point was Claude Code its docs say the advisor read isn't cached, and I can't see a setting to turn it on.

@sermakarevich Matches what I found pricing a month of my own Claude Code logs at API rates. Output (the code it actually writes) was 8% of the bill. Cache reads? 62%. The rereading IS the cost. Short sessions & subagents that hand back a summary are where I'm looking first.