In two days, my AI chief of staff sent out more than 100 jobs. Some came back blocked, and my bill didn't change. It all runs on my $100-a-month Claude Max plan.
The bill part surprised me, so this post has the real numbers, the three setups that failed first, and the five pieces you need to build your own.
One agent to talk to who manages everything. I type a task, it dispatches a worker, and the worker reports back. Getting there took a few wrong turns.
I started with OpenClaw and Hermes. Both were pulling the same Telegram bot token, so messages got dropped and doubled. That was my mistake, not theirs.
Then Hermes had defaulted to Opus, so every message shipped about 120,000 tokens of history with it, and a retry loop fired 16 or more Opus calls in a single turn. My credits were gone by the end of the day. I didn't delete Hermes. It's just not the brain.
Fable 5.1 lists at $10 in and $50 out per million tokens. Opus 5.5 lists at $4 and $20. But the price wasn't the real problem. The context was.
For one week I measured where my money went. The model rereading the whole conversation every turn was about two-thirds of the bill. The words it actually wrote were a tiny slice. The average Opus turn reread about 350,000 tokens to write about 800.
A long session is a bill that grows on every turn, and I was paying it.
Every turn, the model rereads the entire conversation so far. Early in a session that is a small stack of pages. Deep into a session it is a book. Reading about 360,000 tokens to write about 800 means the model is rereading roughly 450 times more than it writes. You are not paying for the answer. You are paying for the rereading. That is why a session that never resets gets more expensive on every turn, and why switching to a "cheaper" model did not fix it for me.
I looked at Paperclip and did research on how other people run agents. I took two ideas and left the products: a hard limit per worker that I control, and a record of who's running.
Opus 5.5 got cheaper in September, and that day I made three moves:
Then two bugs made it real. The relay stalled on a prompt when it tried to exit, and a relay that needs me isn't a relay. And workers died whenever the boss relayed, so I detached them.
The price cut made it affordable. The relay script, the ledger, and the kill timer are what made it stop needing me.
The demo fleet in the video uses dummy workers, not my real setup. You see one place to talk to it. I type a task and it dispatches a worker. When a worker finishes, it appends one line to a shared report file: done or blocked. Blocked is a good outcome, because it means the worker stopped instead of guessing. Last, there's a ledger: who's running, what's queued.
I'm not rounding in my favor. More than 100 jobs went out in two days, and some came back blocked. Cost per turn grows with depth, because the model rereads the whole conversation every turn:
| Turn depth | Cache-read tokens per turn |
|---|---|
| Turns 1 to 10 | about 86,000 |
| Past turn 150 | about 312,000 |
My weekly Opus cache reads went from 837 million to about 2,100 million, up more than 150%. That's because it's doing far more work. But the expensive long turns dropped by half, and my bill didn't change. It all runs on the $100 Max plan.
So is Claude Max worth it? For this use, it is, because the thing that was eating money (long sessions that never reset) was a design problem, not a pricing problem..
Before you blame the plan or the model, look at three things:
If you've been burned by an agent that ran up the bill or dropped your messages, it's usually the same two things: the session never resets, and nothing tells it to stop. Fix those, and the tool matters a lot less.
For the other half of the picture, how to keep an AI from wrecking files while it works, read Claude Code Deleted 48,000 Files.
The Boss Kit is free. It is the rules file my own agent runs on.
Get the Boss KitI send what I'm building, what worked, and what broke. Free, no fluff.
Free. Unsubscribe anytime.I'm a marketer who builds his own tools. More at adamgarceau.com.