I analyzed 187 Claude Code sessions. $6,744 worth of tokens. Here's where they actually went.

I've been using Claude Code heavily for the past month. Building trading bots, automation tools, side projects. I knew I was burning through tokens but never looked at the numbers. So I built a small CLI to parse my local session data. The result: 187 sessions. 3.3 billion tokens. $6,744 equivalent API cost. I'm on Max, so this is equivalent API cost, not what I actually paid. But the token patterns are what matter here. 97% of my tokens were something I couldn't control That was the first surprise. 97% were cache reads. Every turn, Claude re-reads the entire conversation context. Think of it like re-reading an entire book every time you turn a page. The good news: cache reads are cheap ($1.5/M tokens) and completely normal. The bad news: it means the part you can actually control is tiny. Only 2.8% of my tokens were controllable. Of that, 92.5% was cache creation (CLAUDE.md, MCP tools, system prompt loading), 6.6% was Claude's actual output, 0.9% was my input. What I wouldn't have cau