How I used 367 billion tokens in 30 days
I used 367 billion tokens in the last 30 days. My peak day was 37 billion tokens on GPT 5.6 Sol.
The entire workflow
I run one task per cmux workspace with vanilla Codex or Claude Code. Each agent notifies me when it needs a review.
- Press Cmd+Shift+U. cmux jumps to the newest unread notification and focuses the exact workspace and pane.
- Review the dev build, artifact, or bug fix. Reply to the agent.
- Press Cmd+Shift+U again and repeat.
- When no notifications are waiting, create a workspace and start another task.
That loop is the 80/20. cmux keeps active work visible and turns completed agent work into a review queue.
No custom harness
I use vanilla Claude Code and Codex. Their built-in subagents handle the occasional delegated task.
I do not use a custom agent harness or external orchestration. I do not use /loop, and I rarely use /goal.
The economics
At API prices, this usage would have cost about $85,000. We paid $7,400 for 37 ChatGPT Pro 20x accounts and used the quota resets Tibo issued. CodexBar produced the local token and API-cost estimates.
The 367-billion figure still understates the accounts' total usage. Subrouter lets my cofounder share the same account pool, while CodexBar records only the local token usage on my machine.
The workflow keeps my attention on decisions. I review completed work and unblock agents; when the queue clears, I start another task.