For most developers, Claude Code is a black box—you prompt, it codes, magic happens or it doesn't. But one developer decided to crack that box open and document exactly what was happening inside. The results? $514.86 spent over 30 days, 3.8 million tokens consumed, and a side project quietly burning through 63% of the budget.

The Numbers Don't Lie

Tracking every session from April 12 to May 10, 2026 using claudestat—an open-source tool he built specifically for this purpose—the developer logged 50 sessions with some eyebrow-raising metrics. The average efficiency score across all sessions sat at just 69 out of 100. More concerning: the system detected 375 total loops, averaging 7.5 per session. Twelve sessions had more than 10 loops. Nine sessions fell below 50% efficiency.

That One $33 Session

The most expensive single session cost $32.94 on April 26th—a refactoring project where Claude was left running unsupervised for hours. The agent went in circles, editing files that triggered additional edits, creating a cascade of redundant tool calls. Thirty-one loops detected. Efficiency score: 35 out of 100. 'That's more than my daily coffee budget for a month,' the developer noted.

Where the Money Actually Goes

The project breakdown reveals the real story: one hobby project called claudetrace consumed $326.44—63% of total spending. The developer admits he would have guessed closer to 40%. 'This is the kind of insight you only get from tracking,' he wrote. Other projects like wodrival ($61.32), claudestat itself ($60.35), and conductor ($39.40) rounded out the bill.

Loops: The Silent Budget Killer

The most valuable takeaway from this experiment is the loop problem. Most wasteful sessions followed a predictable pattern: Read → Edit → Edit → Edit → Read → Edit → Edit → Edit. Claude attempts something, it doesn't work as expected, so it tries slightly differently and repeats. These loops compound quickly when you're not watching.

Bash Commands Add Up

Perhaps surprisingly, the Bash tool accounted for 35-45% of total costs across all sessions. The developer recommends batching commands rather than running them one at a time—a simple optimization that could save serious money for heavy users.

How to Monitor Your Own Spending

The author built @statforge/claudestat specifically so others could replicate this analysis. Installation is straightforward: npm install -g @statforge/claudestat, then claudestat install and claudestat start before using Claude Code normally. Export data with claudestat export json > my-data.json to analyze costs by project, session efficiency, or loop frequency.

Key Takeaways

  • Loops are expensive—watch for the same tool called 3+ times in a row or 'Let me try...' messages from Claude
  • Set quota alerts at 70%, 85%, and 95% to prevent runaway sessions before they drain your wallet
  • Batch Bash commands aggressively—they're likely eating more of your budget than you realize
  • One project probably dominates your spending; tracking reveals truths that intuition misses

The Bottom Line

The $514 bill isn't necessarily a warning against AI-assisted development—it's a wake-up call about visibility. Without tools like claudestat, developers are flying blind while their tokens burn. If you're serious about using Claude Code at scale, you need the same instrumentation you'd apply to any production system. The black box era is over.