I Tracked My Claude Code Token Spend for a Week. Here's What Actually Surprised Me.

Last week I decided to actually track my Claude Code token usage in real time instead of just checking the dashboard at the end of the month. The results changed how I code. The Setup I used a macOS menu bar tool called TokenBar that sits in your menu bar and shows live token count + dollar cost as you work. It supports OpenAI, Claude, Gemini, Cursor, OpenRouter, and Copilot. The key thing: you see spend while you code, not after. What Surprised Me 1. Context window resets are expensive Every time I started a new Claude Code session without clearing context properly, the first message was eating 8-12k tokens just on the system prompt + file context. Over a day, that added up to roughly 40% of my total spend. Fix: I started being deliberate about when to start fresh sessions vs continue existing ones. Saved about $3/day just from this. 2. Sub-agents stack fast When Claude Code spawns sub-agents for parallel tasks, each one carries its own context. Three sub-agents running simultaneously