The most expensive mistake: $2922 in one session. And it's not her fault

Back in June I wrote about watching Claude and Codex ping-pong while paying for tokens. Promised to tally it up over time. Did. Not by choice: in two days the limits on both accounts hit zero, so I asked the main agent to find out where it all went.

Sixteen sessions running simultaneously. One of them — $2922.55. More than the other fifteen combined.

Disclaimer. I'm on a subscription, these dollars weren't charged to my card — it's the counter in Claude Code's status bar, tokens recalculated at API prices. But limits run out the same way: when they hit zero, all sixteen sessions stop.

What it was doing

Nothing criminal. A client complained: the villa catalog was showing other people's houses on the cards. The agent went to review the cover images. One run — 124 sub-agents and up to 8.9 million tokens, the longest ran for 5 hours and 2 minutes. There were several runs.

The task is closed, deployed, verified in production. Real work. Only all of it ran at the maximum reasoning level. Each of the 124 was thinking as if solving the problem of the millennium, not just matching a picture to an address.

Where the maximum came from

I was sure I had turned it on in one window.

Claude Code has a /effort command. You type /effort xhigh — and it silently saves that to the machine's shared settings file. As the default for all new sessions. Not for this window — for all of them. Per-session — only through the slider and the s key. I didn't know about that. Same goes for the /model command.

Judging by the timestamp, I set it at night, in that very villa window, to speed up one task. From that night on, every new session on the machine started at maximum. Including the one that cuts reels.

The second tier

The video editing agent — $557.45. Opus at maximum was running ffmpeg and placing subtitles. There's nothing to reason about there. The farm agent — $522.92. Opus with a million-token context was reviewing PRs and running tests.

Where did they get Opus? From my own rule. The shared instructions file said: "for subtasks — always opus". That line was copied verbatim into the editing project. The agent wasn't choosing the expensive model. It was following what I had written.

️ What I fixed

Effort level — down to high, model overrides removed. Rewrote the rule: opus — where judgment is needed: architecture, reviews, contested decisions. Sonnet — where it's mechanical: tests, scripts, template-based edits, verification.

Open sessions don't see the change — the setting is read at startup. Until you restart, each session lives with whatever it started with.

And about health. The dollar amount isn't a diagnosis, it's a measure of work volume. The diagnosis is how many times a session compressed its own memory. One of the farm sessions — 282 times over three months. The villa one — 48 times in a week. Five like that are queued for a restart.

To check your own setup — one command: open ~/.claude/settings.json and look at what's set in effortLevel and model. Not in the window. In the file. The window shows what you set for yourself. The file shows what everyone else will get.

I thought I had cranked up one room. I cranked up the whole building. And the counter was running for all of it.